docs: 完成 video workbench 重构案例库

docs: archive video workflow investigations
docs: align video workbench with GPT V2 workflow
2026-06-23 08:46:13 +08:00 · 2026-06-23 08:08:48 +08:00 · 2026-06-23 07:58:44 +08:00
23 changed files with 3153 additions and 39 deletions
--- a/README.md
+++ b/README.md
@ -1,15 +1,15 @@
 # Video Workbench

-Video Workbench is the local execution workspace for dimensional video and slide-style output. It receives GPT planning material from the user, keeps lightweight project records, and produces local assets such as prompts, images, narration audio, slide/page files, and handoff notes.
+Video Workbench is the local execution workspace for dimensional video and slide-style output. It receives GPT V2 planning Markdown from the user, keeps lightweight project records, and produces local execution plans, visual-system assets, prompts, images, narration audio, slide/page files, and handoff notes.

 It is a sibling of Writing Workbench, not a subfolder of it.

 ## Working Model

 ```text
-GPT / ChatGPT      = planning, outlines, storyboards, execution packages
+GPT / ChatGPT      = stage 0-5 planning Markdown, outlines, visual systems, page/shot skeletons
 User               = decisions, source material, accepted planning files
-Video Workbench    = project notes, execution, generated assets, handoff
+Video Workbench    = stage 6+ execution planning, Prompt Advisor work, visual-system materialization, generated assets, handoff
 CCPE System        = local work agents, runtimes, model cards, interfaces
 Skills Vault       = reusable automation skills and deterministic tools
 ```
@ -23,24 +23,34 @@ projects/
  project-id/
    project.md
    intake/
+    execution-plan.md
+    visual-system/
+      visual-system.md
+      refs/
+      iterations/
    slides/
      slides.md
      s01/
      s02/

 handoff/
-tmp/
+docs/
+  cases/
+investigations/
 garden-gpt-image-2/
 requirements/
 ccpe-consumption/
 ```

 - `projects/` is the only project entry point. Project status and series membership are recorded inside project files, not by moving directories.
- `projects/<project>/intake/` holds user-provided GPT planning output, source notes, execution packages, references, and other accepted inputs.
+- `projects/<project>/intake/` holds user-provided GPT V2 stage 0-5 Markdown, source notes, references, and other accepted inputs.
 - `projects/<project>/project.md` is a Codex-maintained project note and file map. It should link to content in `intake/`, not duplicate it.
- `projects/<project>/slides/` is the working area for slide, page, or shot execution. Each `sNN/` folder keeps that unit's prompts, narration, generated images, audio, and local notes together.
+- `projects/<project>/execution-plan.md` is the project-level execution control file for visual work. It records the current strategy, batch, round, and next decision instead of duplicating the project map.
+- `projects/<project>/visual-system/visual-system.md` records the current accepted visual system, materialization targets, iteration log, failure attribution, and next action. Reference images and iteration artifacts stay under `refs/` and `iterations/`.
+- `projects/<project>/slides/` is the working area for slide, page, or shot execution. Each `sNN/` folder keeps that unit's prompts, narration, generated images, audio, and local notes together. Files inside `sNN/` must carry the `sNN` prefix, and iteration files use `sNN-vN-type.ext`, such as `s03-v1-prompt.md`.
 - `handoff/` is for repo-level, cross-project, or context-window handoff notes.
- `tmp/` is for local temporary files.
+- `docs/cases/` is the Codex-side local case-pattern library extracted from old real GPT cases. It preserves execution patterns, not old prompts, JSON packages, paths, or global text-rendering rules.
+- `investigations/` is for research reports, experiments, comparisons, and draft workflow proposals. It replaces `tmp/` for durable investigation evidence.
 - `garden-gpt-image-2/` is reserved for local image-generation prompt/workflow traces when that tool path is used.

 Series indexes can be added later under `projects/_series/` when a real series needs more than project-local notes.
@ -49,6 +59,9 @@ Series indexes can be added later under `projects/_series/` when a real series n

 - Manual editing and publishing remain human-controlled unless explicitly automated later.
 - Video Workbench does not default to full video editing, image-to-video generation, or platform publishing.
+- Codex does not ask GPT V2 for final image prompts, Codex JSON, output paths, task lists, or image generation parameters.
+- `makeup-still/` is not a recommended project directory. Character, scene, template, diagram, and background anchors belong under `visual-system/` when needed.
+- `tmp/` is deprecated. New reusable or reviewable investigation material belongs under `investigations/`.
 - Missing CCPE assets should be requested under `requirements/ccpe/`.
 - Missing reusable automation should be requested under `requirements/skills-vault/`.
 - All local Markdown, JSON, YAML, text, prompt, discussion, and handoff files are UTF-8 by default. PowerShell reads should use explicit `-Encoding UTF8`.
--- a/VIDEO_WORKBENCH.md
+++ b/VIDEO_WORKBENCH.md
@ -4,17 +4,19 @@

 This project is the local execution workspace for dimensional video and slide-style output.

-It turns accepted GPT planning material and user-provided sources into project notes, slide/page execution folders, generated assets, and handoff material. It is optimized for practical continuation, not heavy governance.
+It turns accepted GPT V2 planning Markdown and user-provided sources into project notes, local execution plans, visual-system assets, slide/page execution folders, generated assets, and handoff material. It is optimized for practical continuation, not heavy governance.

 ## 2. Collaboration Roles

 ```text
-GPT / ChatGPT   = plans, outlines, storyboards, execution packages
+GPT / ChatGPT   = stage 0-5 planning Markdown, outlines, visual systems, page/shot skeletons
 User            = decisions, accepted inputs, source/reference material
-Codex           = project notes, file mapping, execution, local assets
+Codex           = stage 6+ execution planning, Prompt Advisor work, visual-system materialization, iteration, local assets
 ```

-Codex should not rewrite GPT planning into duplicate project content. The accepted planning files belong in the project `intake/` directory.
+GPT V2 does not provide final image prompts, Codex JSON, output paths, task lists, or image generation parameters. Codex should not request those from GPT as missing inputs.
+
+Codex should not rewrite GPT planning into duplicate project content. The accepted planning files belong in the project `intake/` directory, and Codex turns them into local execution decisions in `execution-plan.md`.

 ## 3. Project Layout

@ -25,6 +27,11 @@ projects/
  project-id/
    project.md
    intake/
+    execution-plan.md
+    visual-system/
+      visual-system.md
+      refs/
+      iterations/
    slides/
      slides.md
      s01/
@ -33,6 +40,8 @@ projects/

 Use root-level `handoff/` for repository-level, cross-project, series-level, or context-window handoff notes.

+Use `docs/cases/` for Codex-side local case patterns extracted from old real GPT cases. These documents preserve reusable structure, granularity, asset layering, speaker-note style, and review dimensions. They do not restore old GPT final prompts, Codex JSON execution packages, output paths, or global text-rendering rules.
+
 Do not split projects into `active`, `completed`, `abandoned`, `series`, or `standalone` folders. Status and series membership belong in project files because both can change during real production.

 If a series needs more structure, create `projects/_series/` when the first real series needs it.
@ -44,31 +53,82 @@ If a series needs more structure, create `projects/_series/` when the first real
 - project status;
 - medium type;
 - source and reference links;
- which `intake/` files are the outline, storyboard, execution package, or background notes;
+- which `intake/` files are GPT V2 planning Markdown, source notes, reference notes, or background notes;
 - current execution focus;
+- key local execution files;
 - next action.

-It should not duplicate the full outline, storyboard, shot design, or execution package.
+It should not duplicate the full outline, storyboard, page design, shot design, or execution plan.

-## 5. Intake
+## 5. Execution Plan
+
+Visual projects must have:
+
+```text
+projects/<project>/execution-plan.md
+```
+
+This is a living Codex-maintained execution control file. It records:
+
+- the current medium branch;
+- the current execution focus;
+- which visual-system assets need materialization;
+- the current small-batch iteration strategy;
+- which `sNN` units are in the active batch;
+- the current round;
+- the next decision.
+
+It should not repeat the project ID, full intake file index, or project map from `project.md`.
+
+## 6. Intake

 `projects/<project>/intake/` holds material provided or accepted by the user:

- GPT planning documents;
+- GPT V2 stage 0-5 planning Markdown;
 - outlines;
- storyboards;
- detailed slide or shot designs;
- execution packages;
+- storyboards, page skeletons, or shot skeletons;
 - background notes;
 - reference images, audio notes, or source pointers.

 File naming can be agreed during the first real project. Do not create a rigid naming scheme before the real GPT outputs are known.

-## 6. Slides And Shots
+## 7. Visual System
+
+Visual projects use:
+
+```text
+projects/<project>/visual-system/
+  visual-system.md
+  refs/
+  iterations/
+```
+
+`visual-system.md` is the single record for the project's public visual system. It should cover:
+
+- the current accepted visual system;
+- materialization targets;
+- trial prompts and generated reference assets;
+- review notes;
+- failure attribution;
+- next action.
+
+Use optional subfolders only when the project needs them:
+
+```text
+visual-system/characters/
+visual-system/scenes/
+visual-system/templates/
+visual-system/diagrams/
+visual-system/backgrounds/
+```
+
+Do not create a separate `makeup-still/` directory. Character anchors and identity references belong under `visual-system/`.
+
+## 8. Slides And Shots

 Use `slides/` as the execution unit area for both video shots and PPT-style pages.

-`projects/<project>/slides/slides.md` is the lightweight mapping and progress tracker, similar to Writing Workbench's `drafting/drafting.md`.
+`projects/<project>/slides/slides.md` is the unit execution fact table. It records each page or shot's source, batch, current prompt version, current image version, review state, acceptance state, and notes. Project-level iteration strategy belongs in `execution-plan.md`.

 Each `slides/sNN/` folder keeps one slide, page, or shot together:

@ -76,16 +136,28 @@ Each `slides/sNN/` folder keeps one slide, page, or shot together:
 slides/
  slides.md
  s01/
-    prompt files
-    voiceover files
-    generated images
-    generated audio
-    local notes
+    s01-metadata.json
+    s01-unit-spec.md
+    s01-visual-brief.md
+    s01-v1-prompt.md
+    s01-v1-image.png
+    s01-v1-review.md
+    s01-v2-prompt.md
+    s01-v2-image.png
+    s01-v2-review.md
 ```

-Exact file names inside `sNN/` should be chosen when that unit is executed.
+All files inside `sNN/` must carry the `sNN` prefix. Iteration files use `sNN-vN-type.ext`, so files from the same version sort together.

-## 7. Word Budget
+## 9. Investigations
+
+Use `investigations/` for research reports, experiments, comparisons, draft workflow proposals, and other reviewable evidence.
+
+`tmp/` is deprecated and should not receive new durable work. Short-lived scratch material should either be deleted or promoted into `investigations/` when it becomes useful evidence.
+
+`garden-gpt-image-2/` remains reserved for tool-level image prompt or workflow traces when that path is used.
+
+## 10. Word Budget

 Use the default standard:

@ -96,7 +168,7 @@ Use the default standard:

 Adjust only when the user specifies a different narration speed.

-## 8. Asset Policy
+## 11. Asset Policy

 Image and audio generation can be automated when available skills or APIs are appropriate.

@ -104,7 +176,9 @@ Video generation from images remains manual or semi-manual for now because outpu

 Editing and publishing are human-controlled unless explicitly automated later.

-## 9. Capability-Request Mode
+Codex is the default prompt owner for final image prompts in this workspace. It may use `gpt-image-2` Advisor or other image tools, but GPT V2 planning Markdown remains a source specification, not a final prompt.
+
+## 12. Capability-Request Mode

 When a required capability is missing, create a supplier ticket and pause the dependent work.

@ -119,7 +193,7 @@ Use `requirements/ccpe/` for local work agents, CCPE-Lite artifacts, agent specs

 Use `requirements/skills-vault/` for deterministic automation tools, scripts, exporters, batch processors, and installable automation skills.

-## 10. Text Encoding
+## 13. Text Encoding

 All local project documents are UTF-8 unless a file explicitly declares another encoding.

@ -132,7 +206,7 @@ Select-String -Path <path> -Pattern <pattern> -Encoding UTF8

 Do not wait for mojibake or a failed first read before applying UTF-8.

-## 11. CCPE Consumption
+## 14. CCPE Consumption

 Video Workbench consumes CCPE through the single external interface:

--- a/docs/cases/README.md
+++ b/docs/cases/README.md
@ -0,0 +1,65 @@
+# Video Workbench Case Pattern Library
+
+This directory is the local case-pattern library for Codex execution work in Video Workbench.
+
+It is extracted from old GPT case files only at the level of structure, granularity, page or storyboard method, speaker-note style, visual asset layering, and review dimensions. It is not a prompt library and not a copy of the old GPT knowledge base.
+
+## Source Basis
+
+The first two patterns were localized from:
+
+- `knowledge-vault/prompts/GPT/强哥的策划导演/30_CASE_典型视频分镜案例.md`
+- `knowledge-vault/prompts/GPT/强哥的策划导演/31_CASE_典型培训AI_PPT案例.md`
+
+Those source files remain historical GPT assets in Knowledge Vault. Video Workbench uses the extracted patterns below as Codex-side local execution guidance.
+
+## Cases
+
+| case | use when | local output focus |
+| --- | --- | --- |
+| [science-video-page-style-case.md](science-video-page-style-case.md) | Turning a deep article, model point, or public-facing concern into a PPT-style science video | Shot/page granularity, narration pacing, visual metaphor, review criteria |
+| [training-ai-ppt-case.md](training-ai-ppt-case.md) | Turning AI education, method training, product enablement, or workshop content into a teachable slide deck | Teaching unit design, speaker notes, interaction, editable slide structure |
+
+## Local Use
+
+Use these cases after GPT V2 stage 0-5 planning has been accepted into a project `intake/` directory. Codex then converts the accepted planning into:
+
+- `project.md` for project map and current focus;
+- `execution-plan.md` for execution strategy, batch, round, and next decision;
+- `visual-system/visual-system.md` for visual-system materialization and review;
+- `slides/slides.md` and `slides/sNN/` for page or shot execution facts and assets.
+
+## What To Extract
+
+Extract these parts from old or future real cases:
+
+- medium branch and audience;
+- logic-to-page or logic-to-shot decomposition;
+- page or shot function;
+- one-unit-one-purpose granularity;
+- editable page text vs visual asset vs narration or speaker notes;
+- visual-system anchors and reusable motifs;
+- review dimensions and acceptance criteria;
+- batch/iteration advice.
+
+## What Not To Inherit
+
+Do not inherit these legacy assumptions:
+
+- GPT outputs final image prompts.
+- GPT outputs Codex JSON execution packages.
+- GPT specifies local output paths, task lists, or generation parameters.
+- the old assumption that every image prompt must forbid readable text.
+- old source filenames or schemas become current Video Workbench contracts.
+
+For current Video Workbench, text placement is an execution decision. Body copy is usually rendered in the editable PPT or video page layer; image-generated labels, formulas, diagrams, or text are allowed only when the execution plan and review criteria explicitly require that layer.
+
+## Review Dimensions
+
+When adding a new case pattern, check:
+
+- Does the pattern state which medium branch it serves?
+- Does it describe unit granularity without copying a full old case?
+- Does it separate source logic, page or shot copy, visual assets, and narration/speaker notes?
+- Does it provide review criteria that Codex can apply during small-batch iteration?
+- Does it avoid old Prompt, JSON, path, and global text-rendering rules?
--- a/docs/cases/science-video-page-style-case.md
+++ b/docs/cases/science-video-page-style-case.md
@ -0,0 +1,156 @@
+# Science Video Page-Style Case Pattern
+
+This pattern is for PPT-style science videos: a deep article or model is reduced into a short public-facing video made of page-like shots, narration, and visual assets.
+
+It is extracted from the old `30_CASE` video storyboard case. The original topic, prompts, Codex JSON, output paths, and global text prohibitions are not carried forward.
+
+## Use When
+
+Use this pattern when the target is:
+
+- a 3-10 minute public-facing science, cognition, AI, product-thinking, or method-explainer video;
+- a PPT-style video with page/shot images plus narration;
+- a video that explains a mechanism, risk, contradiction, method, or model through metaphors and staged visual pages;
+- not a drama short, not a full video-editing workflow, and not a character-continuity MV.
+
+## Core Pattern
+
+```text
+source point
+-> public-facing concern
+-> controlling metaphor
+-> video outline
+-> shot/page list
+-> shot/page deep spec
+-> narration
+-> visual asset brief
+-> review and iteration
+```
+
+The case value is not the original topic. The reusable method is "page-style explanation": each shot behaves like a compact visual argument, not like a literal filmed scene.
+
+## Granularity
+
+Default unit size:
+
+- 6-10 units for a short explainer;
+- 25-70 seconds per unit;
+- about 90-260 Chinese characters per unit;
+- one unit carries one logic function;
+- one high-density method unit may be longer, but it still needs one clear governing frame.
+
+Common unit functions:
+
+| function | purpose |
+| --- | --- |
+| hook | create urgency or recognisable public concern |
+| story anchor | give the viewer a concrete person, situation, or conflict |
+| controlling metaphor | make the whole mechanism memorable |
+| mechanism reveal | explain the hidden process or failure mode |
+| practical frame | turn the concept into methods, checks, or choices |
+| closing elevation | return to the larger judgment or changed mental model |
+
+## Local Unit Spec
+
+For Video Workbench, put this in `slides/sNN/sNN-unit-spec.md` or use it as the basis for that file:
+
+| field | purpose |
+| --- | --- |
+| `unit_id` | `sNN` identifier used by local files |
+| `source_anchor` | source paragraph, GPT V2 stage-5 row, or accepted intake note |
+| `narrative_function` | hook, story anchor, metaphor, mechanism, method, or closing |
+| `core_message` | one sentence the viewer should retain |
+| `page_copy` | short overlay text or subtitle-layer copy, if needed |
+| `voiceover_intent` | what narration must explain beyond page copy |
+| `visual_task` | what the visual asset must make visible |
+| `asset_layers` | background, metaphor object, human element, diagram, overlay, audio |
+| `continuity_link` | how this unit connects to the previous and next unit |
+| `review_focus` | the two or three things to inspect first in iteration |
+
+This is a local execution spec, not a final image prompt. Final prompt files remain Codex-owned `sNN-vN-prompt.md` artifacts when image generation is actually performed.
+
+## Page And Asset Layering
+
+Separate layers before prompting or generating images:
+
+| layer | owns |
+| --- | --- |
+| source logic | the concept, contradiction, risk, or method from accepted input |
+| page text | the few words that may be rendered by PPT/video page layer |
+| visual asset | metaphor scene, background, figures, diagram base, mood, composition |
+| narration | full explanation, transition, examples, and punchline |
+| review notes | whether the visual asset really carries the logic |
+
+Do not treat "image contains no readable text" as a global rule. Decide per unit:
+
+- If text must be editable or frequently revised, keep it in the PPT/video page layer.
+- If labels, formulas, UI snippets, or diagram words are essential to the visual explanation, allow them only with explicit review criteria.
+- If the image model is likely to garble precise wording, keep precise wording outside the generated image.
+
+## Shot/Page Design Method
+
+For each unit, answer in this order:
+
+1. What is the viewer supposed to understand or feel at this moment?
+2. What is the simplest visible metaphor or situation that makes it concrete?
+3. What page text is useful, if any?
+4. What must narration explain that the image should not carry alone?
+5. Which part of the unit is high risk: logic, metaphor, composition, text placement, or visual generation?
+6. What would make the unit fail review?
+
+The visual should not decorate the narration. It should do at least one of these jobs:
+
+- compress an abstract mechanism;
+- expose a contradiction;
+- make a risk feel real;
+- separate similar concepts;
+- hold attention while narration explains a dense idea;
+- signal a transition in the argument.
+
+## Review Dimensions
+
+Review generated or drafted units against these dimensions:
+
+| dimension | pass condition |
+| --- | --- |
+| logic clarity | one unit maps to one clear logic point |
+| metaphor fit | the metaphor explains the point instead of becoming a side joke |
+| public accessibility | a non-specialist can understand the visible situation |
+| narration fit | voiceover adds explanation, not redundant caption reading |
+| composition | image leaves safe space for page text if page text is planned |
+| asset feasibility | the image can be generated or assembled without fragile exact text |
+| continuity | the unit advances from the previous one and sets up the next one |
+| medium fit | it remains a science/explainer page, not an accidental drama scene |
+
+## Small-Batch Strategy
+
+Do not generate every unit first. Pick representative high-risk units:
+
+- one hook or opening unit;
+- one controlling-metaphor unit;
+- one mechanism-explanation unit;
+- one practical-method or dense information unit;
+- one closing or elevation unit when tone is uncertain.
+
+Record the selected batch in `execution-plan.md`, then track unit-level facts in `slides/slides.md`.
+
+## Abstracted Example Shape
+
+```md
+## s03 Unit Spec
+
+- narrative_function: controlling metaphor
+- core_message: The tool is useful only when matched to the weight of the real-world problem.
+- page_copy: two short lines, rendered outside the generated image unless the execution plan says otherwise
+- voiceover_intent: contrast everyday low-risk use with high-stakes decision use
+- visual_task: make the mismatch between lightweight confidence and large external risk visible
+- asset_layers:
+  - background: large-scale risk environment
+  - metaphor object: small helpful tool that looks insufficient at scale
+  - human element: ordinary user facing the environment
+  - overlay: optional page text layer
+- review_focus:
+  - risk scale is obvious
+  - metaphor does not become comic noise
+  - composition supports 16:9 page use
+```
--- a/docs/cases/training-ai-ppt-case.md
+++ b/docs/cases/training-ai-ppt-case.md
@ -0,0 +1,162 @@
+# Training AI PPT Case Pattern
+
+This pattern is for training, AI education, customer enablement, internal workshops, and method lectures that need a teachable slide deck.
+
+It is extracted from the old `31_CASE` training AI PPT case. The original topic, prompts, Codex JSON, output paths, and global text prohibitions are not carried forward.
+
+## Use When
+
+Use this pattern when the target is:
+
+- a training deck, course deck, lecture deck, workshop deck, or AI education PPT;
+- a slide sequence that must support live explanation, pause, discussion, and later reuse;
+- a deck where each page is a teaching unit, not a video shot;
+- not a customer proposal whose primary goal is decision conversion.
+
+## Core Difference
+
+Training PPT pages are not scenes. They are teachable units.
+
+Each page should answer:
+
+- What should the learner understand after this page?
+- What misconception or resistance might exist before this page?
+- What example, analogy, contrast, diagram, or exercise lowers the learning barrier?
+- What can the learner do, judge, or remember after this page?
+
+## Learning Path
+
+Common sequence:
+
+```text
+problem entry
+-> concept or model
+-> method frame
+-> example or contrast
+-> practice or migration
+-> summary and action
+```
+
+Not every deck needs all six parts, but the page order should behave like a learning path rather than a list of impressive points.
+
+## Page Granularity
+
+Default rule:
+
+- one page solves one teaching goal;
+- slide copy stays sparse and editable;
+- speaker notes carry explanation, examples, and transition;
+- interaction is included when the page benefits from learner reflection;
+- abstract models need a diagram, matrix, loop, ladder, map, or other visible structure.
+
+Avoid one page trying to explain a concept, prove a case, give a method, run an exercise, and conclude the section at the same time.
+
+## Local Page Spec
+
+For Video Workbench, put this in `slides/sNN/sNN-unit-spec.md` or use it as the basis for that file:
+
+| field | purpose |
+| --- | --- |
+| `unit_id` | `sNN` identifier used by local files |
+| `slide_role` | opening, concept, model, contrast, case, practice, summary, transition |
+| `teaching_goal` | what the learner should understand, change, or be able to do |
+| `core_message` | one sentence that gives the page its point |
+| `slide_copy` | editable title, subtitle, bullets, labels, or quoted line |
+| `layout` | comparison, triangle, matrix, timeline, loop, ladder, map, flow, dashboard |
+| `visual_asset_brief` | background, metaphor, concept visual, diagram base, or scene asset |
+| `speaker_notes` | how the instructor explains the page |
+| `interaction` | question, quick vote, reflection, mini exercise, or none |
+| `transition` | how this page connects to adjacent pages |
+| `review_focus` | what Codex should inspect during iteration |
+
+This is a local execution spec, not a final image prompt or JSON execution package.
+
+## Slide And Asset Layering
+
+Keep these layers separate:
+
+| layer | owns |
+| --- | --- |
+| editable slide text | titles, subtitles, bullet points, labels, formulas, exact wording |
+| layout | structure that controls learner attention |
+| visual asset | background, metaphor scene, illustration, diagram base, icons, texture |
+| speaker notes | explanation, example, pacing, transition, instructor emphasis |
+| interaction | question or exercise that turns listening into retrieval or judgment |
+
+Generated images should not be asked to carry precise deck body copy by default. If the design needs diagram labels or visible text inside an image, document that choice in `execution-plan.md` or the unit review criteria.
+
+## Speaker Notes Style
+
+Speaker notes should do three jobs:
+
+- explain the key concept in plain language;
+- give a concrete life, classroom, business, or product example;
+- bridge to the next page.
+
+Good notes are not page copy repeated aloud. They let the slide stay sparse while still giving the instructor enough material to speak for 1-2 minutes when needed.
+
+## Common Page Patterns
+
+| pattern | use when | layout hint |
+| --- | --- | --- |
+| structural problem | establish why the topic matters | triangle, tension map, before/after |
+| misconception correction | replace a shallow belief with a better frame | split screen, false/true contrast |
+| model introduction | name and explain a reusable model | matrix, loop, layered diagram |
+| method frame | turn concept into steps or checks | flow, ladder, checklist, flywheel |
+| case comparison | show how AI changes a task or judgment | traditional vs AI-enabled |
+| practice page | make learners apply the frame | prompt, scenario, quick exercise |
+| summary chain | close the logic path | timeline, chain, staircase, map |
+
+"Traditional vs AI-enabled" is especially useful for AI training case pages, but it is a pattern choice, not a global rule.
+
+## Review Dimensions
+
+Review drafted or generated pages against these dimensions:
+
+| dimension | pass condition |
+| --- | --- |
+| teaching goal | one page has one clear learner outcome |
+| copy density | slide text is sparse enough to scan and edit |
+| layout clarity | structure helps understanding rather than decorating the page |
+| speaker notes | notes explain, exemplify, and transition |
+| asset layering | visual assets do not replace editable slide text by accident |
+| interaction | any question or exercise has a clear teaching reason |
+| sequence | the page advances the learning path |
+| live usability | an instructor can pause on the page and teach from it |
+
+## Small-Batch Strategy
+
+Do not build the full deck visually before validating the teaching system. Pick representative pages:
+
+- one opening or problem page;
+- one abstract model page;
+- one method-frame page;
+- one case-comparison page;
+- one practice or summary page if interaction or closing tone is uncertain.
+
+Record the selected batch in `execution-plan.md`, then track page-level status in `slides/slides.md`.
+
+## Abstracted Example Shape
+
+```md
+## s08 Unit Spec
+
+- slide_role: model introduction
+- teaching_goal: Learners understand that the model is a decision aid, not a decorative framework.
+- core_message: A usable model changes what people can notice, compare, and improve.
+- slide_copy:
+  - title: short model name
+  - subtitle: one-line use claim
+  - bullets: three editable labels or dimensions
+- layout: center matrix with one highlighted region
+- visual_asset_brief: clean diagram base with subtle learning-path motion; exact labels remain editable in the slide layer
+- speaker_notes:
+  - define the model in plain language
+  - give one concrete classroom or work example
+  - explain why the next page moves from model to use case
+- interaction: ask learners which dimension they currently under-observe
+- review_focus:
+  - model structure is legible
+  - page is not crowded
+  - notes can support live explanation
+```
--- a/docs/context-handoff.md
+++ b/docs/context-handoff.md
@ -19,12 +19,14 @@ Video Workbench keeps local project notes, execution outputs, and handoff record
 ## Collaboration Contract

 ```text
-GPT / ChatGPT   = planning and execution-package authoring
+GPT / ChatGPT   = stage 0-5 planning Markdown, visual-system design, page/shot skeletons
 User            = decisions and accepted input placement
-Codex           = file mapping, project notes, execution, generated assets
+Codex           = stage 6+ execution planning, Prompt Advisor work, visual-system materialization, iteration, generated assets
 ```

-Most content files should come from GPT output accepted by the user and placed in `projects/<project>/intake/`. Codex-maintained files should be short continuation aids and execution records.
+Most planning files should come from GPT V2 output accepted by the user and placed in `projects/<project>/intake/`. Codex-maintained files should be local execution aids and execution records.
+
+Codex should not ask GPT V2 for final image prompts, Codex JSON, output paths, task lists, or image generation parameters. GPT V2 planning Markdown is a source specification, not an execution command.

 ## Project Shape

@ -44,6 +46,11 @@ Recommended project-local structure:
 projects/<project-id>/
  project.md
  intake/
+  execution-plan.md
+  visual-system/
+    visual-system.md
+    refs/
+    iterations/
  slides/
    slides.md
    s01/
@ -52,15 +59,29 @@ projects/<project-id>/

 Use root `handoff/` for repo-level, cross-project, series-level, or context-window handoff notes.

+Use `docs/cases/` as the local Codex case-pattern library. It is allowed to extract structure, page or shot granularity, speaker-note style, visual asset layering, and review dimensions from old GPT cases, but it must not re-import old final-prompt, Codex JSON, output-path, or global text-rendering rules.
+
 ## Project Files

-`project.md` is a Codex-maintained small note. It should link to intake files instead of duplicating the outline, storyboard, or execution package.
+`project.md` is a Codex-maintained small note. It should link to intake files and key local execution files instead of duplicating the outline, storyboard, page design, or execution plan.

-`intake/` holds GPT planning material, user-provided source material, references, execution packages, and other accepted input files.
+`intake/` holds GPT V2 stage 0-5 planning Markdown, user-provided source material, references, and other accepted input files.

-`slides/slides.md` is the slide/page/shot mapping and progress tracker. It is the Video Workbench equivalent of a lightweight drafting tracker.
+`execution-plan.md` is the visual project's local execution control file. It records the current medium branch, visual-system materialization targets, small-batch iteration strategy, active `sNN` units, current round, and next decision. It does not repeat the project map from `project.md`.

-`slides/sNN/` keeps all files for one slide, page, or shot together: prompts, narration, generated images, generated audio, and local notes.
+`visual-system/visual-system.md` records the current accepted visual system, materialization targets, iteration log, failure attribution, and next action. Character anchors, scene anchors, templates, diagrams, and backgrounds belong under `visual-system/`, not a separate `makeup-still/`.
+
+`slides/slides.md` is the slide/page/shot execution fact table. It records each unit's source, batch, prompt version, image version, review state, acceptance state, and notes. It does not replace `execution-plan.md`.
+
+`slides/sNN/` keeps all files for one slide, page, or shot together: prompts, narration, generated images, generated audio, review notes, and local notes. Every file inside `sNN/` must carry the `sNN` prefix; iteration files use `sNN-vN-type.ext`, such as `s03-v1-prompt.md`.
+
+## Investigations
+
+Use `investigations/` for research reports, experiments, comparisons, draft workflow proposals, and reviewable evidence.
+
+`tmp/` is deprecated. New durable investigation material should go under `investigations/`; short-lived scratch material should be deleted when no longer needed.
+
+`garden-gpt-image-2/` remains reserved for tool-level image prompt or workflow traces when that path is used.

 ## Text Encoding

--- a/investigations/2026-06-23-gpt-image-2-advisor-prompt-quality-v2-broader-use-cases/01-分工与目录改造草案.md
+++ b/investigations/2026-06-23-gpt-image-2-advisor-prompt-quality-v2-broader-use-cases/01-分工与目录改造草案.md
@ -0,0 +1,414 @@
+# 01-分工与目录改造草案
+
+```yaml
+status: draft
+date: 2026-06-23
+scope:
+  - video-workbench 本地执行体系
+  - GPT「强哥的策划导演」协作边界
+  - gpt-image-2 Advisor 模式的提示词职责
+premise:
+  - v2 提示词质量调研结论已被接受
+  - 当前只确认提示词质量，不确认最终出图质量
+non_goals:
+  - 不直接修改 GPT Builder 或 GPT 知识库
+  - 不直接新增 video-workbench/docs 下的稳定文档
+  - 不验证生图效果、图生图效果或视频生成效果
+  - 不迁移既有项目目录
+```
+
+## 一、前提结论
+
+这次改造的起点不是“人物定妆照”本身，而是一个更上层的职责判断：
+
+> GPT 负责方向、结构、教学或叙事意图；Codex 在本地负责把视觉资产需求转成可执行、可追踪、可复用的最终图像提示词。
+
+v2 调研已经支持这个判断。尤其在高频场景里，`gpt-image-2` Advisor 模式比 Codex 直接写提示词更稳定：
+
+- 科普视频 / 分镜类内容：Advisor 更擅长把抽象概念拆成可视化结构、画面层级、动态感、非文字依赖表达。
+- 培训 AI / PPT 类内容：Advisor 更擅长区分页面文字、讲解逻辑、视觉资产、标题安全区和留白。
+- 人物造型 / 定妆照：Advisor 仍有价值，但在当前使用场景中权重最低，只作为条件能力沉淀。
+
+因此后续改造不应围绕“GPT 如何写更好的生图 Prompt”展开，而应围绕“GPT 如何把视觉任务交接给 Codex，由 Codex 本地生成最终 Prompt”展开。
+
+## 二、新的三方分工
+
+| 角色 | 主要职责 | 不再默认承担 |
+| --- | --- | --- |
+| 强哥 | 提供目标、素材、判断标准；决定采用哪版方向、哪版提示词、哪版资产 | 不需要手工维护每个视觉资产的最终 Prompt 版本 |
+| GPT「强哥的策划导演」 | 做选题理解、结构规划、分镜 / 页面设计、教学意图、视觉方向、导演卡、验收标准、Codex 交接包 | 不默认输出最终生图 Prompt / 图生图 Prompt / 镜头 Prompt |
+| Codex / video-workbench | 本地建档、拆解任务、调用 `gpt-image-2` Advisor 模式生成最终提示词、保存提示词和资产版本、组织复盘 | 不替代强哥做方向决策；不把调研结论直接当稳定规范 |
+| `gpt-image-2` Skill | 作为 Codex 本地提示词工程能力，提供 Advisor 模式和结构化图像提示词方法 | 不是 GPT 的知识库文件，也不是 GPT 必须直接掌握的完整执行流程 |
+
+一句话边界：
+
+> GPT 输出“视觉任务说明书”；Codex 输出“可执行图像提示词”；生成结果和版本记录留在 video-workbench 项目目录。
+
+## 三、媒介输出边界
+
+### 1. 科普视频 / 视频分镜
+
+这是当前最高权重场景。
+
+GPT 应输出：
+
+- 主题理解和核心解释路径
+- 每页 / 每镜的讲解目标
+- 画面构图方向
+- 视觉隐喻或演示机制
+- 页面文字建议
+- 可交给 Codex 的导演卡
+- 该页 / 该镜的验收标准
+
+Codex 应输出：
+
+- `image-prompt-vN.md`
+- 必要时的 `image-edit-prompt-vN.md`
+- 生成或外部生成后的图片资产
+- 对照验收标准的复盘记录
+
+### 2. 培训 AI / PPT 类内容
+
+这是第二高权重场景，且比人物类更常见。
+
+GPT 应重点处理：
+
+- 教学目标
+- 页面信息层级
+- 页面文字
+- 讲解顺序
+- 视觉资产用途
+- 适合 PPT 的留白和标题区
+
+Codex 应重点处理：
+
+- 视觉资产提示词
+- 图像里不承载长段文字的约束
+- 页面文字与图片文字的边界
+- 素材版本和复用记录
+
+原则：
+
+> PPT / 网页 / 视频页面上的文字属于页面层，不应默认塞进图片提示词里。图片提示词主要负责视觉资产，而不是负责排版整页 PPT。
+
+### 3. 人物造型 / 定妆照
+
+这是条件场景，不是本次体系的中心。
+
+只在项目中出现稳定人物、主持人、角色、连续剧情或需要跨镜头形象一致时启用。
+
+GPT 可输出：
+
+- 人物身份定位
+- 角色气质
+- 服装 / 发型 / 妆造方向
+- 表情和姿态参考
+- 连续使用注意点
+
+Codex 应输出并保存：
+
+- 定妆照提示词
+- 生成后的定妆图
+- 人物一致性参考说明
+- 后续镜头调用该人物时的引用记录
+
+## 四、目录改造草案
+
+### 1. `investigations/`
+
+用途：调研、验证、草案和过渡文件。
+
+本文件目前就放在 `investigations/` 下，因为它还不是稳定工作规范。只有强哥确认后，才把其中的稳定部分迁移到 `docs/`。
+
+建议规则：
+
+- 调研报告、实验记录、对比样例、草案放这里。
+- 不把这里的文件默认当作项目运行规范。
+- 被接受的结论要迁移到 `docs/` 或项目模板后才算稳定知识。
+
+### 2. `docs/`
+
+用途：video-workbench 的稳定本地知识库。
+
+本次草案通过后，建议新增以下文档：
+
+```text
+docs/image-prompt-advisor-workflow.md
+docs/visual-asset-prompt-rubric.md
+docs/page-text-vs-image-prompt-boundary.md
+docs/makeup-still-directory-policy.md
+```
+
+四个文档的优先级：
+
+1. `image-prompt-advisor-workflow.md`
+   - 规定 Codex 如何从 GPT 导演卡进入 Advisor 模式。
+   - 规定输出哪些提示词文件、如何命名、如何复盘。
+
+2. `visual-asset-prompt-rubric.md`
+   - 固化 v2 调研使用的提示词质量维度。
+   - 用于判断提示词是否足够具体、可执行、可生成。
+
+3. `page-text-vs-image-prompt-boundary.md`
+   - 规定页面文字、PPT 文字、视频字幕、图片内文字的边界。
+   - 避免把可编辑文字错误地塞进图片生成任务。
+
+4. `makeup-still-directory-policy.md`
+   - 规定什么时候需要 `makeup-still/`。
+   - 规定人物提示词、定妆图、人物一致性说明如何保存。
+
+### 3. `projects/<project-id>/intake/`
+
+用途：保存已接受的 GPT 上游输出和项目输入。
+
+建议保存：
+
+```text
+projects/<project-id>/intake/source.md
+projects/<project-id>/intake/gpt-direction.md
+projects/<project-id>/intake/gpt-outline.md
+projects/<project-id>/intake/gpt-execution-package.md
+projects/<project-id>/intake/references.md
+```
+
+其中 `gpt-execution-package.md` 可以包含 GPT 输出的导演卡集合、页面目标、视觉方向、验收标准和 Codex 交接说明。
+
+### 4. `projects/<project-id>/slides/sNN/`
+
+用途：保存单页 / 单镜头的执行资产。
+
+当前 workbench 已经使用 `slides/sNN/` 作为单位目录，因此本次不另起 `shots/` 体系。即使是视频分镜，也先沿用 `slides/sNN/` 作为“一屏 / 一镜 / 一页”的统一执行单元。
+
+建议结构：
+
+```text
+projects/<project-id>/slides/s01/
+  director-card.md
+  visual-brief.md
+  image-prompt-v1.md
+  image-prompt-v2.md
+  image-v1.png
+  image-v2.png
+  review-notes.md
+  voiceover.md
+  metadata.json
+```
+
+文件边界：
+
+- `director-card.md`：来自 GPT 或由 Codex 整理的导演卡，描述该页 / 该镜要达成什么。
+- `visual-brief.md`：面向视觉资产的需求说明，可由 GPT 输出，也可由 Codex 从上游材料压缩而来。
+- `image-prompt-vN.md`：Codex 调用 Advisor 模式后的最终图像提示词。
+- `image-vN.png`：对应提示词生成或外部平台生成后的图片。
+- `review-notes.md`：强哥 / Codex 对该页资产的验收和修改记录。
+- `metadata.json`：可选，用于记录模型、平台、版本、来源路径和引用关系。
+
+### 5. `projects/<project-id>/makeup-still/`
+
+用途：人物定妆、角色形象和跨镜头人物一致性。
+
+这是条件目录，不是所有项目都需要。
+
+启用条件：
+
+- 项目有固定主持人、角色或拟人 IP。
+- 多张图片 / 多个镜头需要保持同一人物形象。
+- 定妆图会被后续图生图、视频生成、封面或分镜反复引用。
+
+建议结构：
+
+```text
+projects/<project-id>/makeup-still/
+  identity-brief.md
+  identity-prompt-v1.md
+  identity-ref-v1.png
+  wardrobe-prompt-v1.md
+  wardrobe-ref-v1.png
+  pose-ref-notes.md
+  consistency-notes.md
+  acceptance-notes.md
+```
+
+规则：
+
+- 定妆提示词由 Codex 生成。
+- 定妆图可以由 Codex 生成，也可以由强哥在外部平台生成后放入该目录。
+- 无论图片从哪里来，只要成为项目人物参考，就应进入 `makeup-still/`。
+- 不出现人物或人物不需要连续一致时，不创建该目录。
+
+### 6. `garden-gpt-image-2/`
+
+用途：工具级 prompt / workflow trace，不是项目资产的最终归档地。
+
+建议规则：
+
+- Skill 调试、通用样例、工具行为记录可以放这里。
+- 某个项目真正采用的提示词和图片，应复制或记录到对应 `projects/<project-id>/slides/sNN/` 或 `makeup-still/`。
+- 不让项目资产只停留在 `garden-gpt-image-2/`。
+
+### 7. `tmp/`
+
+用途：短期临时文件。
+
+建议规则：
+
+- 临时测试可以放这里。
+- 调研证据、可复用草案、已接受结论不放这里。
+- 能被后续复盘引用的内容，应放入 `investigations/`。
+
+### 8. `requirements/ccpe/` 与 `requirements/skills-vault/`
+
+用途：能力缺口外部请求。
+
+建议规则：
+
+- 如果是 GPT / CCRA 知识库改造需求，先形成改动清单或审阅包，不在 video-workbench 内伪装成已完成改造。
+- 如果是可复用 Skill、脚本、校验器、打包器能力，写入 `requirements/skills-vault/`。
+- 如果涉及 CCPE 正式参与者、审查协议或跨 Agent 调度，写入 `requirements/ccpe/`。
+
+## 五、GPT 知识库 / CCRA 改动清单
+
+这部分只记录“需要改什么”，不在本文件中执行。
+
+### 1. `强哥的策划导演.md`
+
+需要把默认输出从“最终 Prompt”降级为“导演层交接物”。
+
+建议调整：
+
+- 将“画面 Prompt / 镜头 Prompt / 生图 Prompt / 图生图 Prompt”改为条件输出。
+- 默认输出改为：
+  - 视觉资产意图
+  - 页面 / 分镜导演卡
+  - 视觉 brief
+  - 可交给 Codex 的执行说明
+  - 验收标准
+- 增加边界说明：最终图像提示词由 Codex 结合 `gpt-image-2` Advisor 模式在本地生成。
+
+### 2. `00_CORE_附录B_内容类型识别与分流.md`
+
+需要把科普视频、培训 PPT、MV / 人物类任务分流到不同视觉交接级别。
+
+建议新增或调整字段：
+
+```text
+prompt_owner: codex
+gpt_output_level: director_card | visual_brief | execution_package
+advisor_mode_required: true | false
+```
+
+### 3. `00_CORE_附录C_视频_MV_短剧_深度执行模式.md`
+
+需要拆分“导演层”和“最终提示词层”。
+
+建议调整：
+
+- GPT 输出镜头意图、构图、节奏、视觉隐喻、验收标准。
+- 不默认输出最终 image prompt。
+- 人物连续性场景增加 `makeup-still/` 交接提醒。
+
+### 4. `00_CORE_附录E_Codex_交接与本地执行.md`
+
+这是最需要调整的核心文件之一。
+
+建议新增：
+
+- Codex 接收 GPT 导演卡后，进入 `gpt-image-2` Advisor 模式。
+- Codex 在项目目录保存 `image-prompt-vN.md`。
+- Codex 保存生成图片、外部图片来源和验收记录。
+- Codex 将人物定妆资产保存到 `makeup-still/`。
+
+### 5. `21_SYSTEM_多工具协同与流程编排.md`
+
+需要把“GPT 直接产 Prompt”改成“GPT 编排任务，Codex 负责本地提示词执行”。
+
+建议强调：
+
+- GPT 是策划导演，不是本地资产管理员。
+- Codex 是执行、归档、版本管理和提示词工程落地者。
+- Advisor 模式是 Codex 的本地能力，不是 GPT 必须内置的完整流程。
+
+### 6. `30_CASE_典型视频分镜案例.md`
+
+保留作为科普视频高频样例，但需要改标注。
+
+建议：
+
+- 将旧版直接 Prompt 示例标为 legacy baseline 或“提示词颗粒度参考”。
+- 新增一版“GPT 应输出的导演卡 / visual brief”。
+- 新增一版“Codex 应生成的最终 image prompt”作为本地执行示例。
+
+### 7. `31_CASE_典型培训AI_PPT案例.md`
+
+保留并提高权重，因为它更接近实际高频使用。
+
+建议：
+
+- 明确页面文字由 PPT / 网页层承载。
+- 图片提示词只负责视觉资产。
+- GPT 输出教学目标、页面结构、视觉资产用途。
+- Codex 输出视觉资产提示词和版本记录。
+
+### 8. `10_STYLE_强哥偏好与审美取向.md`
+
+可暂缓，只在需要统一美学偏好时再改。
+
+建议：
+
+- 不把风格偏好扩写成复杂 prompt 规范。
+- 只保留可影响导演卡和 visual brief 的偏好描述。
+
+### 9. `90_ARCHIVE/`
+
+不作为本轮改造对象。
+
+建议：
+
+- 只读参考。
+- 不上传为 GPT 当前知识。
+- 不把 archive 里的旧 prompt 体系当作当前职责边界。
+
+## 六、video-workbench 稳定文档候选
+
+本草案通过后，建议在 `docs/` 下新增四个稳定文档。
+
+| 文档 | 解决的问题 | 来源 |
+| --- | --- | --- |
+| `docs/image-prompt-advisor-workflow.md` | Codex 如何从 GPT 导演卡生成最终图像提示词 | 本草案 + v2 调研 |
+| `docs/visual-asset-prompt-rubric.md` | 如何评价提示词质量 | v2 调研评分维度 |
+| `docs/page-text-vs-image-prompt-boundary.md` | 页面文字和图片提示词的职责边界 | 培训 AI / PPT 高频场景 |
+| `docs/makeup-still-directory-policy.md` | 什么时候创建人物定妆目录 | 条件人物场景 |
+
+这些文档应在强哥确认本草案后再创建，避免把草案过早固化。
+
+## 七、建议执行顺序
+
+1. 强哥审阅并修订本草案。
+2. Codex 将被接受的部分迁移为 `docs/` 下的稳定文档。
+3. Codex 准备 GPT / CCRA 知识库改动清单或补丁包。
+4. 用一个真实的科普视频或培训 AI PPT 项目做单页 / 单镜试运行。
+5. 根据试运行结果，再决定是否需要模板、脚本或校验器。
+
+## 八、验收标准
+
+本草案被接受时，应满足：
+
+- 高频权重正确：科普视频 > 培训 AI / PPT > 人物造型。
+- GPT 不再默认承担最终图像提示词生成。
+- Codex 明确承担本地图像提示词工程和文件归档。
+- `makeup-still/` 是条件目录，不是所有项目的默认目录。
+- `investigations/` 与 `docs/` 的边界清楚。
+- 没有把提示词质量结论夸大为最终图片质量结论。
+- 没有把 GPT 知识库改造伪装成已经执行完成。
+
+## 九、暂不做
+
+本轮不做：
+
+- 不直接生图。
+- 不验证最终图片质量。
+- 不改 GPT Builder。
+- 不改 knowledge-vault 中的 GPT 知识库文件。
+- 不新增稳定 `docs/` 文档。
+- 不迁移现有项目目录。
+- 不引入自动化脚本。
--- a/investigations/2026-06-23-gpt-image-2-advisor-prompt-quality-v2-broader-use-cases/01-分工与目录改造草案2.0.md
+++ b/investigations/2026-06-23-gpt-image-2-advisor-prompt-quality-v2-broader-use-cases/01-分工与目录改造草案2.0.md
@ -0,0 +1,597 @@
+# 01-分工与目录改造草案2.0
+
+```yaml
+status: draft
+date: 2026-06-23
+scope:
+  - video-workbench 本地执行流程
+  - GPT「强哥的策划导演 V2」阶段0-5 Markdown 接收方式
+  - Codex 阶段6以后执行、试错、提示词和资产落盘规则
+basis:
+  - GPT V2 已完成改造
+  - GPT 不再输出最终图像 Prompt
+  - GPT 不再输出 Codex JSON 或执行包
+  - Codex 从 GPT V2 Markdown 策划稿中自行生成本地执行计划
+non_goals:
+  - 不修改 GPT V2 知识库
+  - 不恢复旧版 Codex JSON 流程
+  - 不要求一次性生成全套 Prompt
+  - 不设置默认导出或剪辑流程
+```
+
+## 一、总体边界
+
+GPT「强哥的策划导演 V2」的默认终点是阶段5：
+
+```text
+阶段0：任务路由
+阶段1：原始势能扫描
+阶段2：降维方向提案
+阶段3：媒介总纲设计
+阶段4：通用表达与视觉系统设计
+阶段5：分镜 / 页面 / 内容骨架
+```
+
+video-workbench 的默认起点是阶段6：
+
+```text
+阶段6：本地执行启动
+```
+
+Codex 不再向 GPT 索要：
+
+- 最终图像 Prompt；
+- Negative Prompt；
+- Codex JSON；
+- video-workbench task list；
+- output_path；
+- 图像生成参数；
+- 批量出图任务单。
+
+Codex 接收的是 GPT V2 的阶段0-5 Markdown 策划结果，并在本地完成执行计划、视觉系统物化、提示词生成、资产生成、review、失败归因和迭代。
+
+## 二、角色分工
+
+| 角色 | 负责 | 不负责 |
+| --- | --- | --- |
+| 强哥 | 建立或授权具体项目；确认 GPT 策划结束；决定是否进入 Codex 执行；审阅关键成果 | 不维护每个资产的提示词版本和迭代记录 |
+| GPT V2 | 阶段0-5策划；总纲；通用表达与视觉系统；页面 / 分镜 / 内容骨架 | 不输出最终 Prompt；不输出 Codex JSON；不规定本地目录和执行路径 |
+| Codex / video-workbench | 阶段6以后本地执行；解析 GPT Markdown；生成 `execution-plan.md`；物化 `visual-system/`；生成和迭代单元资产 | 不把 GPT 策划稿当成执行命令；不等待 GPT 提供 JSON |
+| `gpt-image-2` Advisor | 为 Codex 提供图像提示词工程支持 | 不替代项目级执行计划和本地 review |
+
+一句话边界：
+
+> GPT V2 给“策划骨架和公共系统”；Codex 把它变成“本地执行计划、视觉系统、提示词、资产和迭代记录”。
+
+## 三、本地流程2.0
+
+本地流程压缩为五个阶段。
+
+### 1. Intake Locked
+
+触发条件：
+
+- 具体项目已经建立或由强哥授权建立；
+- `projects/<project-id>/intake/` 已放入 GPT V2 给出的全部 Markdown 文件；
+- 强哥明确表示 GPT 策划工作结束，可以进入 Codex 执行。
+
+Codex 在这一阶段做：
+
+- 确认 `intake/` 中有哪些 GPT V2 文件和用户素材；
+- 创建或更新 `project.md`；
+- 补齐本地执行需要的目录和索引文件；
+- 不把 `intake/` 中的 GPT Markdown 改写成另一套完整内容；
+- 不回头要求 GPT 提供 Prompt、JSON 或 task list。
+
+建议项目起始结构：
+
+```text
+projects/<project-id>/
+  project.md
+  intake/
+  execution-plan.md
+  visual-system/
+    visual-system.md
+  slides/
+    slides.md
+```
+
+如果项目不是视觉类任务，可以不创建 `visual-system/` 和 `slides/`。
+
+### 2. Build Local Execution Plan
+
+`parse stage outputs` 不作为单独文档阶段。Codex 可以在讨论和分析中解析 GPT V2 Markdown，但最终判断要汇总进项目级 `execution-plan.md`。
+
+视觉类项目默认必须有：
+
+```text
+projects/<project-id>/execution-plan.md
+```
+
+`execution-plan.md` 不是旧 GPT JSON 的替代品，也不是一次性批量任务清单。它是本地执行的活文档，记录当前项目如何试、怎么分批、跑到哪一轮、下一步做什么。
+
+建议内容：
+
+```md
+# Execution Plan
+
+项目索引：见 `project.md`。
+
+## 1. Execution Scope
+- medium_branch:
+- current_phase:
+- current_focus:
+- next_decision:
+
+## 2. Visual-System Materialization Plan
+- assets_to_fix:
+- first_round_targets:
+- acceptance_criteria:
+- open_questions:
+
+## 3. Unit Iteration Plan
+| batch | units | purpose | entry_condition | status | decision |
+
+## 4. Current Round
+- round:
+- units_or_assets:
+- prompts_to_generate:
+- images_to_generate:
+- review_focus:
+- next_action:
+
+## 5. Decision Log
+| date | decision | reason | affected files |
+```
+
+使用规则：
+
+- 每次进入新一轮视觉系统试跑或页面 / 镜头试跑前，先更新 `execution-plan.md`。
+- 它记录“执行策略、批次、轮次、当前决策”。
+- 它可以包含进度概览，但不替代 `slides/slides.md` 的单元执行明细。
+- 它可以先小批次推进，不要求全项目 Prompt 一次性生成完。
+- 项目 ID、媒介类型、`intake/` 文件索引和关键路径由 `project.md` 维护，不在 `execution-plan.md` 中重复展开。
+
+### 3. Materialize Visual System
+
+这是阶段6以后最关键的本地执行阶段。
+
+统一使用：
+
+```text
+projects/<project-id>/visual-system/
+```
+
+不再使用独立的：
+
+```text
+projects/<project-id>/makeup-still/
+```
+
+人物定妆、角色一致性、场景参考、PPT 背景、页面模板、图示风格、公共提示词、图片母版和参考资产，都归入 `visual-system/`。
+
+建议结构：
+
+```text
+projects/<project-id>/visual-system/
+  visual-system.md
+  refs/
+  iterations/
+```
+
+可按项目需要增加子目录：
+
+```text
+visual-system/
+  characters/
+  scenes/
+  templates/
+  diagrams/
+  backgrounds/
+  refs/
+  iterations/
+```
+
+文件边界：
+
+- `visual-system.md`：三合一记录公共视觉系统的当前采用版本、物化目标、迭代日志、失败归因和下一步。
+- `refs/`：外部参考图、用户提供素材、GPT 阶段4提到的参考需求。
+- `iterations/`：保存每轮试跑的提示词、图片、review 和修改判断。
+
+不同媒介的物化重点：
+
+| 类型 | `visual-system/` 重点 |
+| --- | --- |
+| 培训 / 科普 AI PPT | 封面模板、章节页模板、概念页模板、案例页模板、图示风格、公式 / 图表 / 文字承载策略 |
+| 客户方案 PPT | 商务视觉基调、页面密度、客户决策路径图示、价值证明图示、品牌和留白规则 |
+| 科普视频 / PPT式视频 | 公共画面基调、统摄性隐喻、视觉母题、图示风格、页面文字 / 字幕 / 图表承载策略 |
+| MV / 剧情 / 人物连续性 | 人物身份锚点、服装锚点、姿态锚点、场景锚点、道具锚点、光线和情绪色彩 |
+
+注意：
+
+- `generate reference assets` 并入本阶段。
+- `review and iterate visual system` 并入本阶段。
+- 视觉系统可以迭代多轮，不要求第一轮固定。
+- 只有 `visual-system.md` 的“当前接受版本”中记录的内容，才算项目级公共视觉系统已接受版本。
+
+### 4. Generate And Iterate Units
+
+这一阶段不采用“一次性生成全套 Prompt，再批量执行”的方式。
+
+默认策略是小批次试错、小批次迭代：
+
+```text
+先选 1 / 5 / 7 等代表性页面或镜头
+→ 生成这些单元的提示词和图片
+→ review
+→ 需要则进入二稿
+→ 通过后再选 2 / 4 / 8 等下一批
+→ 最后处理剩余单元
+```
+
+批次策略写在：
+
+```text
+projects/<project-id>/execution-plan.md
+```
+
+单元执行事实写在：
+
+```text
+projects/<project-id>/slides/slides.md
+```
+
+两者分工：
+
+| 文件 | 记录什么 | 不记录什么 |
+| --- | --- | --- |
+| `execution-plan.md` | 当前试错策略、批次安排、轮次状态、下一步判断 | 不展开每个单元的完整执行文件 |
+| `slides/slides.md` | 每个页面 / 镜头的具体执行状态 | 不替代项目级试错策略 |
+| `slides/sNN/` | 单个页面 / 镜头的提示词、图片、review、metadata | 不保存其他单元的资产 |
+
+`slides/slides.md` 建议记录：
+
+```md
+| unit | source | batch | spec | prompt | image | review | accepted | notes |
+| --- | --- | --- | --- | --- | --- | --- | --- | --- |
+| s01 | GPT stage5 row 1 | batch-01 | done | v2 | v2 | pass | yes | uses visual-system v1 |
+| s05 | GPT stage5 row 5 | batch-01 | done | v1 | v1 | revise | no | composition too dense |
+```
+
+每个 `slides/sNN/` 目录的文件名必须带 `sNN` 前缀。
+
+示例：
+
+```text
+projects/<project-id>/slides/s03/
+  s03-metadata.json
+  s03-unit-spec.md
+  s03-visual-brief.md
+  s03-v1-prompt.md
+  s03-v1-image.png
+  s03-v1-review.md
+  s03-v2-prompt.md
+  s03-v2-image.png
+  s03-v2-review.md
+  s03-final-notes.md
+```
+
+规则：
+
+- 不使用无前缀的 `image-prompt.md`、`review-notes.md`、`metadata.json` 作为默认命名。
+- 多个单元同时打开时，文件名本身必须能识别属于哪个页面或镜头。
+- 迭代文件采用 `sNN-vN-type.ext`，让同一版本的 prompt、image、review 排在一起。
+- 如果同一单元有多张资产，也继续带前缀和版本，例如 `s03-v1-diagram.png`、`s03-v1-background.png`。
+- `sNN-unit-spec.md` 来自 GPT 阶段5和本地执行计划，不是最终 Prompt。
+- `sNN-vN-prompt.md` 才是 Codex 生成的最终图像提示词。
+
+### 5. Closeout When Needed
+
+默认不设置独立的：
+
+```text
+export_final_assets
+write_project_handoff
+```
+
+平时续作依赖：
+
+- `project.md`
+- `execution-plan.md`
+- `visual-system/visual-system.md`
+- `slides/slides.md`
+- `slides/sNN/` 内文件
+
+只有以下情况才写 handoff：
+
+- 跨线程继续；
+- 长时间暂停；
+- 项目交付给其他工具或人；
+- 需要形成阶段性归档；
+- 强哥明确要求。
+
+handoff 可放在项目内或 repo 级 `handoff/`，按实际上下文决定。
+
+## 四、媒介分支流程
+
+### 1. 培训 / 科普 AI PPT
+
+默认流程：
+
+```text
+Intake Locked
+→ execution-plan.md
+→ visual-system/ 固定页面模板和图示系统
+→ 代表页小批次试跑
+→ slides/sNN/ 逐页生成和迭代
+```
+
+重点：
+
+- 页面模板先于全量页面生成；
+- 代表页应覆盖概念页、案例页、练习页、图表页；
+- 文字、公式、图表、标签可以由 PPT 层、图像层或后期层承载，由 Codex 在执行计划中判断；
+- 不设置“图片不能有文字”的全局规则。
+
+### 2. 客户方案 PPT
+
+默认流程：
+
+```text
+Intake Locked
+→ execution-plan.md
+→ visual-system/ 固定商务质感、页面密度和图表规则
+→ 决策关键页小批次试跑
+→ slides/sNN/ 逐页生成和迭代
+```
+
+重点：
+
+- 先试封面、客户痛点页、关键判断页、方案结构页、价值证明页；
+- 视觉资产必须服务客户决策；
+- 不把 PPT 页当成视频镜头。
+
+### 3. 科普视频 / PPT式视频
+
+默认流程：
+
+```text
+Intake Locked
+→ execution-plan.md
+→ visual-system/ 固定公共画面基调、视觉母题和图示风格
+→ 认知关键镜头小批次试跑
+→ slides/sNN/ 逐镜头生成和迭代
+```
+
+重点：
+
+- 先试开场、核心机制解释、反转、总结升维等关键镜头；
+- 口播和画面功能要匹配；
+- 画面不是口播装饰；
+- 是否出现文字、公式、图表，由解释任务决定。
+
+### 4. MV / 剧情 / 人物连续性
+
+默认流程：
+
+```text
+Intake Locked
+→ execution-plan.md
+→ visual-system/ 固定人物、服装、姿态、场景、道具和情绪色彩
+→ 高风险镜头小批次试跑
+→ slides/sNN/ 逐镜头生成和迭代
+```
+
+重点：
+
+- `makeup-still/` 作废，不再作为独立目录；
+- 人物定妆只是 `visual-system/characters/` 的一部分；
+- 先固定多锚点参考资产，再做镜头；
+- 高风险镜头包括人物近景、全身、动作、跨场景、情绪变化、多人关系。
+
+### 5. 文章 / 口播 / 纯文本
+
+默认不走完整视觉流程。
+
+如果没有图片、PPT、视频页面、MV 或视觉资产需求，可以不创建：
+
+- `visual-system/`
+- `slides/`
+
+如果后续要转成视频、PPT 或音频可视化，再进入对应视觉流程。
+
+## 五、目录2.0
+
+推荐项目结构：
+
+```text
+projects/<project-id>/
+  project.md
+  intake/
+  execution-plan.md
+  visual-system/
+    visual-system.md
+    refs/
+    iterations/
+  slides/
+    slides.md
+    s01/
+      s01-metadata.json
+      s01-unit-spec.md
+      s01-visual-brief.md
+      s01-v1-prompt.md
+      s01-v1-image.png
+      s01-v1-review.md
+    s02/
+```
+
+条件目录：
+
+```text
+visual-system/characters/
+visual-system/scenes/
+visual-system/templates/
+visual-system/diagrams/
+visual-system/backgrounds/
+```
+
+不再推荐：
+
+```text
+projects/<project-id>/makeup-still/
+```
+
+## 六、文件职责
+
+### `project.md`
+
+轻量项目地图。
+
+记录：
+
+- 项目状态；
+- 媒介类型；
+- `intake/` 文件列表；
+- 当前执行焦点；
+- 关键文件链接；
+- 下一步。
+
+不记录完整执行计划，不复制 GPT 策划稿。
+
+### `execution-plan.md`
+
+项目级执行活文档。
+
+记录：
+
+- Codex 如何解析 GPT V2 阶段0-5；
+- 本项目属于哪个媒介分支；
+- 阶段4要先物化哪些公共视觉系统；
+- 当前采用什么小批次试错策略；
+- 当前跑到第几轮；
+- 哪些单元进入当前批次；
+- 下一步做什么；
+- 哪些判断已被强哥接受。
+
+它是旧 GPT JSON 消失后的本地核心控制文件。
+
+### `visual-system/visual-system.md`
+
+公共视觉系统的当前状态、物化计划和迭代记录。
+
+记录：
+
+- 当前已接受的公共视觉系统；
+- 需要物化的视觉目标；
+- 每轮试跑了什么；
+- 使用了哪些提示词；
+- 生成了哪些参考资产；
+- review 发现了什么问题；
+- 失败归因是什么；
+- 哪些版本被接受；
+- 当前公共视觉系统使用哪个版本。
+
+### `slides/slides.md`
+
+页面 / 镜头执行明细表。
+
+记录：
+
+- 每个 `sNN` 的来源；
+- 属于哪个批次；
+- 当前 prompt 版本；
+- 当前图片版本；
+- review 状态；
+- 是否通过；
+- 备注。
+
+不承担项目级批次策略。
+
+### `slides/sNN/`
+
+单个页面 / 镜头资产目录。
+
+所有文件名必须带 `sNN` 前缀。
+
+## 七、稳定文档后续改造
+
+本草案确认后，应再更新 video-workbench 的稳定文档。
+
+优先改：
+
+```text
+README.md
+VIDEO_WORKBENCH.md
+docs/context-handoff.md
+```
+
+重点改动：
+
+- 把 `GPT = execution packages` 改为 `GPT V2 = 阶段0-5策划 Markdown`；
+- 把 `Codex = execution` 扩展为 `Codex = 阶段6以后本地执行、Prompt Advisor、视觉系统物化、资产迭代和落盘`；
+- 去掉对 GPT JSON、Prompt、路径任务单的期待；
+- 增加 `execution-plan.md`；
+- 增加 `visual-system/visual-system.md`；
+- 明确 `makeup-still/` 不再作为推荐目录；
+- 明确 `slides/sNN/` 文件名前缀和版本命名规则；
+- 增加 `docs/cases/`，承接旧 `30_CASE` / `31_CASE` 的本地化案例模式。
+
+可新增稳定文档：
+
+```text
+docs/gpt-v2-local-execution-workflow.md
+docs/visual-system-materialization.md
+docs/unit-iteration-and-progress.md
+docs/cases/README.md
+docs/cases/science-video-page-style-case.md
+docs/cases/training-ai-ppt-case.md
+```
+
+是否新增这些文档，等本草案确认后再决定。
+
+`docs/cases/` 的定位不是复制旧 GPT 知识库，而是从旧真实案例中抽取 Codex 本地执行可用的模式：
+
+- 科普视频的页面化表达颗粒度；
+- 培训 PPT 的教学单元设计；
+- 页面文字、视觉资产、讲师备注的分层；
+- 分镜 / 页面 review 的判断维度。
+
+不继承：
+
+- GPT 直接输出最终图像 Prompt；
+- GPT 输出 Codex JSON；
+- 旧路径和旧执行包字段；
+- `no readable text` 全局规则。
+
+## 八、硬规则
+
+1. Codex 不向 GPT 索要最终 Prompt。
+2. Codex 不向 GPT 索要 Codex JSON。
+3. Codex 不把 GPT V2 Markdown 当成执行命令。
+4. `execution-plan.md` 是视觉类项目默认必备文件。
+5. `visual-system/visual-system.md` 是公共视觉系统的状态、计划和迭代记录统一文件。
+6. `makeup-still/` 作废，不再作为推荐目录。
+7. 不要求一次性生成全套页面 / 镜头 Prompt。
+8. 默认采用小批次试错、小批次迭代。
+9. `slides/slides.md` 记录单元执行事实，不替代项目级执行计划。
+10. `slides/sNN/` 下所有文件名必须带 `sNN` 前缀，迭代文件采用 `sNN-vN-type.ext`。
+11. 不把 `no readable text` 当成全局图片规则。
+12. 是否由图片承载文字、公式、标签、图表，由 Codex 根据媒介和内容目标在执行层判断。
+13. `export_final_assets` 和 `write_project_handoff` 按需执行，不作为默认阶段。
+14. 旧 `30_CASE` / `31_CASE` 只能抽取为 `docs/cases/` 的案例模式，不作为旧 Prompt / JSON 流程回流入口。
+
+## 九、验收标准
+
+本草案被接受时，应满足：
+
+- 能从 GPT V2 阶段0-5 Markdown 自然进入 Codex 本地执行；
+- 不依赖 GPT Prompt、JSON 或路径任务单；
+- `execution-plan.md` 的职责清楚；
+- `visual-system/` 的职责清楚；
+- `makeup-still/` 已从推荐目录中移除；
+- 支持小批次试错，而不是全量 Prompt 先行；
+- `slides/slides.md` 和 `slides/sNN/` 的职责分离清楚；
+- `slides/sNN/` 文件名前缀和版本命名规则明确；
+- 旧案例进入 `docs/cases/` 的定位清楚；
+- PPT、科普视频、MV / 剧情、纯文本任务有合理分支；
+- 没有把 GPT 给出的参考流程机械照搬成本地流程。
--- a/investigations/2026-06-23-gpt-image-2-advisor-prompt-quality-v2-broader-use-cases/README.md
+++ b/investigations/2026-06-23-gpt-image-2-advisor-prompt-quality-v2-broader-use-cases/README.md
@ -0,0 +1,46 @@
+# GPT Image 2 Advisor Prompt Quality Investigation V2
+
+## Why V2 Exists
+
+The first investigation over-weighted character / MV identity scenarios. That made the result too narrow for the user's actual video-workbench usage.
+
+This V2 uses source scenes from:
+
+- `C:\Users\wangq\Documents\Codex\knowledge-vault\prompts\GPT\强哥的策划导演\30_CASE_典型视频分镜案例.md`
+- `C:\Users\wangq\Documents\Codex\knowledge-vault\prompts\GPT\强哥的策划导演\31_CASE_典型培训AI_PPT案例.md`
+
+The weighted priority is:
+
+```text
+典型视频分镜(科普) > 典型培训AI > 人物造型
+```
+
+## Test Scope
+
+This is only a prompt-quality investigation. It does not generate images and does not claim final image quality.
+
+The current environment is still treated as `gpt-image-2` Advisor / Mode C for this investigation: prompt generation and comparison only.
+
+## Case Mix
+
+| Group | Weight | Cases |
+| --- | ---: | --- |
+| Science video storyboard | 55% | S-01 AI wave, S-03 AI float vs ocean, S-05 logic sandcastle, S-07 three brakes |
+| Training AI PPT | 35% | P8 prompt engineering vs cognitive engineering, P12 score vs growth-state assessment, P18 education AI logic chain |
+| Character / makeup-still | 10% | lightweight directory-boundary sample only |
+
+## Files
+
+- `cases.md`: source-derived cases and evaluation intent.
+- `outputs/direct-prompts.md`: baseline prompts from or modeled closely on the GPT case files.
+- `outputs/advisor-prompts.md`: Advisor-style prompts using `gpt-image-2` structure and relevant templates.
+- `提示词质量综合调研报告.md`: weighted comparison and conclusion.
+
+## Template References Used
+
+- `gpt-image-2/SKILL.md`
+- `references/prompt-writing.md`
+- `references/slides-and-visual-docs/dense-explainer-slides.md`
+- `references/slides-and-visual-docs/educational-diagram-slide.md`
+- `references/infographics/comparison-infographic.md`
+- `references/scenes-and-illustrations/concept-scene.md`
--- a/investigations/2026-06-23-gpt-image-2-advisor-prompt-quality-v2-broader-use-cases/cases.md
+++ b/investigations/2026-06-23-gpt-image-2-advisor-prompt-quality-v2-broader-use-cases/cases.md
@ -0,0 +1,127 @@
+# Cases
+
+## Source Principle
+
+The cases are taken from the two user-specified GPT knowledge files. They are not newly invented story scenarios.
+
+## Group A: Science Video Storyboard, Weight 55%
+
+### V-S01: 被浪拍下水的人
+
+Source: `30_CASE_典型视频分镜案例.md`, S-01.
+
+Intent:
+
+- Opening hook for a science / commentary video.
+- Establish the metaphor that ordinary people are being pushed into the AI wave.
+- Page text will be added outside the generated image.
+
+Prompt risk:
+
+- AI symbols become literal text.
+- The scene becomes a disaster movie instead of a metaphorical visual essay.
+- The composition leaves no title-safe area.
+
+### V-S03: 游泳圈与太平洋
+
+Source: `30_CASE_典型视频分镜案例.md`, S-03.
+
+Intent:
+
+- Governing metaphor: AI float vs the Pacific Ocean.
+- Show scale contrast: tool is light, reality is vast and heavy.
+
+Prompt risk:
+
+- The ring becomes too cute or too central.
+- Ocean scale does not dominate.
+- The metaphor turns into a travel poster.
+
+### V-S05: 逻辑沙堡
+
+Source: `30_CASE_典型视频分镜案例.md`, S-05.
+
+Intent:
+
+- Risk visualization: beautiful AI-generated business logic may collapse under reality.
+- Sandcastle must look precise and fragile; the wave must feel imminent.
+
+Prompt risk:
+
+- Business-plan elements become readable text.
+- The sandcastle looks like a fairy-tale castle rather than fragile logic.
+- The wave lacks pressure.
+
+### V-S07: 防忽悠三板斧
+
+Source: `30_CASE_典型视频分镜案例.md`, S-07.
+
+Intent:
+
+- Practical framework: three brakes before using AI for high-stakes action.
+- Visualize calculation, red-team attack, and stop-loss rule as three gates / brakes.
+
+Prompt risk:
+
+- The three mechanisms are unclear.
+- Text is rendered into the image.
+- The image becomes a generic car safety ad.
+
+## Group B: Training AI PPT, Weight 35%
+
+### T-P8: 提示词工程已死？认知工程才刚开始
+
+Source: `31_CASE_典型培训AI_PPT案例.md`, P8.
+
+Intent:
+
+- Training slide contrast: old prompt tricks vs cognitive engineering.
+- Left/right comparison visual; final page text handled by PPT layer.
+
+Prompt risk:
+
+- The image tries to render the page text.
+- The contrast is too decorative and not teachable.
+- "Cognitive core" becomes a generic sci-fi brain.
+
+### T-P12: 从分数评价到成长状态评价
+
+Source: `31_CASE_典型培训AI_PPT案例.md`, P12.
+
+Intent:
+
+- Training slide contrast: result scoring vs growth-state assessment.
+- Explain a move from static score sheets to dynamic learning path / feedback maps.
+
+Prompt risk:
+
+- The left side becomes unreadable score-sheet text.
+- The right side becomes generic data visualization.
+- The visual loses the teacher / student / AI assessment triangle.
+
+### T-P18: AI教育变革的逻辑链
+
+Source: `31_CASE_典型培训AI_PPT案例.md`, P18.
+
+Intent:
+
+- Summary slide: one logic chain from AI ecosystem to cognitive model to CCPE / SPGM to assessment innovation and human-AI education ecology.
+
+Prompt risk:
+
+- Five nodes lose their sequence.
+- The image renders fake labels.
+- The chain looks like generic "AI transformation" marketing art.
+
+## Group C: Character / Makeup Still, Weight 10%
+
+### C-01: 人物形象目录边界样本
+
+Intent:
+
+- Keep one lightweight character sample only to evaluate whether Advisor prompts support the proposed `makeup-still/` directory.
+
+Prompt risk:
+
+- Overgeneralizing character workflow into all video-workbench work.
+- Mistaking character prompt quality for the whole system's quality.
--- a/investigations/2026-06-23-gpt-image-2-advisor-prompt-quality-v2-broader-use-cases/outputs/advisor-prompts.md
+++ b/investigations/2026-06-23-gpt-image-2-advisor-prompt-quality-v2-broader-use-cases/outputs/advisor-prompts.md
@ -0,0 +1,374 @@
+# Advisor Prompts
+
+These prompts apply `gpt-image-2` Advisor structure: type, goal, layout, subject / scene, style, constraints, and production-use notes.
+
+## V-S01 Advisor: 被浪拍下水的人
+
+```json
+{
+  "type": "science video metaphor still / opening hook",
+  "goal": "Create a 16:9 visual asset for a PPT-style science video opening. The image must carry the metaphor 'ordinary people are being pushed into the AI wave'; final Chinese page text will be added by the video/PPT layer.",
+  "source_scene": {
+    "scene_id": "S-01",
+    "visual_role": "opening_hook",
+    "core_metaphor": "AI wave pushes ordinary people into the water"
+  },
+  "composition": {
+    "aspect_ratio": "16:9",
+    "foreground": "several ordinary people being hit by a huge dark-blue wave; at least one person still holding a phone",
+    "midground": "one bright inflatable ring as the only light playful object, clearly small compared with the wave",
+    "background": "towering ocean wave mixed with faint AI-era visual fragments",
+    "title_safe_area": "leave upper-left or lower-left negative space for two lines of overlay text",
+    "camera": "wide cinematic editorial angle, people small enough to show pressure from the environment"
+  },
+  "visual_symbols": {
+    "ai_fragments": "subtle non-readable UI thumbnails, chat-window silhouettes, image-generation thumbnails, code-like light fragments",
+    "symbol_rule": "symbols should be recognizable by shape, not by readable words"
+  },
+  "style": {
+    "rendering": "cinematic editorial illustration, metaphorical visual essay",
+    "palette": "dark blue ocean, black-blue shadows, small warm amber highlights",
+    "lighting": "dramatic but clean, not horror, not action-movie spectacle"
+  },
+  "constraints": {
+    "must_keep": [
+      "wave pressure is the dominant visual force",
+      "people feel like ordinary users, not heroes",
+      "AI elements are subtle background fragments",
+      "composition supports page text overlay"
+    ],
+    "avoid": [
+      "readable text, logos, watermarks",
+      "literal AI word labels on objects",
+      "monster-wave disaster movie style",
+      "too many UI fragments that make the image cluttered",
+      "cheerful tech marketing mood"
+    ]
+  }
+}
+```
+
+## V-S03 Advisor: 游泳圈与太平洋
+
+```json
+{
+  "type": "science video governing metaphor still",
+  "goal": "Create a 16:9 visual asset that explains the governing metaphor: a light AI float is not enough for the vast Pacific of reality.",
+  "source_scene": {
+    "scene_id": "S-03",
+    "visual_role": "governing_metaphor",
+    "core_metaphor": "tool is light, reality is vast and heavy"
+  },
+  "composition": {
+    "aspect_ratio": "16:9",
+    "subject_scale": "the human figure must be small; ocean and sky occupy at least 75% of the frame",
+    "foreground": "shallow water edge, a small human holding a bright inflatable ring",
+    "midground": "boundary line from shallow water to dark deep water",
+    "background": "vast dark ocean under storm clouds, horizon feels far and overwhelming",
+    "camera": "wide low-distance composition, strong scale contrast",
+    "title_safe_area": "keep one clean area for overlay text"
+  },
+  "style": {
+    "rendering": "cinematic metaphorical illustration",
+    "palette": "dark blue, blue-black ocean, gray storm sky, one controlled bright ring accent",
+    "mood": "sobering, not heroic, not inspirational travel"
+  },
+  "constraints": {
+    "must_keep": [
+      "human must look tiny",
+      "ocean must feel immense and risky",
+      "float must look light, almost absurd",
+      "image must serve the governing metaphor"
+    ],
+    "avoid": [
+      "travel poster beauty",
+      "cute beach vacation",
+      "person centered as hero",
+      "readable text, logos, watermarks",
+      "oversaturated ring that dominates the frame"
+    ]
+  }
+}
+```
+
+## V-S05 Advisor: 逻辑沙堡
+
+```json
+{
+  "type": "science video risk visualization still",
+  "goal": "Create a 16:9 metaphor image: a beautiful AI-generated business plan may be only a fragile logic sandcastle before reality hits.",
+  "source_scene": {
+    "scene_id": "S-05",
+    "visual_role": "risk_visualization",
+    "core_metaphor": "beautiful logical plan vs reality wave"
+  },
+  "composition": {
+    "aspect_ratio": "16:9",
+    "foreground": "a highly detailed sandcastle built like a fragile business architecture: towers, bridges, tiny pathways, neatly arranged but clearly made of sand",
+    "supporting_objects": "scattered abstract chart pages and planning diagrams near the sandcastle, only visual marks, no readable text",
+    "background": "large dark-blue wave curling toward the sandcastle, close enough to feel imminent",
+    "camera": "three-quarter low angle so the sandcastle looks impressive but vulnerable",
+    "title_safe_area": "leave clean upper area or sky area for overlay text"
+  },
+  "style": {
+    "rendering": "metaphorical cinematic editorial illustration",
+    "palette": "warm sand / gold foreground against dark blue wave",
+    "lighting": "beautiful foreground light interrupted by threatening cool shadow from the wave"
+  },
+  "constraints": {
+    "must_keep": [
+      "sandcastle is precise, complex, and fragile",
+      "wave has imminent destructive pressure",
+      "contrast between beautiful logic and harsh reality is obvious",
+      "business-plan symbols remain non-readable"
+    ],
+    "avoid": [
+      "fairy-tale castle made of stone",
+      "calm decorative beach",
+      "readable charts or documents",
+      "wave too far away",
+      "cartoon comedy tone"
+    ]
+  }
+}
+```
+
+## V-S07 Advisor: 防忽悠三板斧
+
+```json
+{
+  "type": "science video practical framework still",
+  "goal": "Create a 16:9 educational metaphor image showing three brakes before high-stakes AI action.",
+  "source_scene": {
+    "scene_id": "S-07",
+    "visual_role": "practical_framework",
+    "framework_items": ["risk calculation", "red-team attack", "stop-loss rule"]
+  },
+  "composition": {
+    "aspect_ratio": "16:9",
+    "main_scene": "a small vehicle rushing toward a cliff edge that drops into a deep dark ocean",
+    "safety_structure": "three clearly separated glowing brake gates between the vehicle and danger",
+    "gate_1": "calculator / ledger icon for risk calculation",
+    "gate_2": "cyber watchdog / red-team icon for attack and critique",
+    "gate_3": "cold contract / stop sign / signed paper icon for stop-loss rule",
+    "flow": "left-to-right or foreground-to-background progression: vehicle -> three gates -> danger",
+    "title_safe_area": "keep a clear overlay area for two lines of page text"
+  },
+  "style": {
+    "rendering": "cinematic educational illustration, clean symbolic structure",
+    "palette": "dark blue risk zone, warm amber gate light, neutral vehicle",
+    "mood": "tense but practical, clear, not cartoonish"
+  },
+  "constraints": {
+    "must_keep": [
+      "three mechanisms are visually distinct",
+      "danger ahead is clear",
+      "gates visibly stop the vehicle",
+      "icons communicate without readable words"
+    ],
+    "avoid": [
+      "readable Chinese or English labels",
+      "generic traffic safety poster",
+      "overly cute watchdog",
+      "more than three gates",
+      "cluttered UI elements"
+    ]
+  }
+}
+```
+
+## T-P8 Advisor: 提示词工程已死？认知工程才刚开始
+
+```json
+{
+  "type": "training PPT comparison visual asset",
+  "goal": "Create a 16:9 visual asset for a training slide comparing old prompt tricks with cognitive engineering. The PPT layer will render all final readable text.",
+  "source_page": {
+    "slide_id": "P8",
+    "teaching_goal": "Show that low-level prompt tricks are aging, while structured expert cognition becomes more important."
+  },
+  "layout": {
+    "page_type": "two-column comparison",
+    "left_column": {
+      "role": "old prompt engineering",
+      "visual": "scattered prompt cards, broken magic-scroll metaphor, loose snippets and role-play masks fading out",
+      "mood": "fragmented, clever but shallow"
+    },
+    "right_column": {
+      "role": "cognitive engineering",
+      "visual": "structured cognitive core / knowledge engine with process loops, evaluation nodes, domain layers, feedback arrows",
+      "mood": "ordered, durable, expert-like"
+    },
+    "transition": "central arrow or transformation bridge from left to right",
+    "title_safe_area": "top band left empty enough for slide title"
+  },
+  "style": {
+    "rendering": "clean academic training slide illustration",
+    "palette": "muted gray on left, deep blue and warm gold on right, off-white background",
+    "information_density": "medium-high, but not a final text slide"
+  },
+  "constraints": {
+    "must_keep": [
+      "clear left/right contrast",
+      "visual logic supports later overlay labels",
+      "cognitive core must show structure: loops, layers, standards",
+      "no final readable text generated in image"
+    ],
+    "avoid": [
+      "generic sci-fi brain without teaching meaning",
+      "too much fake text",
+      "magic fantasy dominates the left side",
+      "marketing-poster hero composition",
+      "logos or watermarks"
+    ]
+  }
+}
+```
+
+## T-P12 Advisor: 从分数评价到成长状态评价
+
+```json
+{
+  "type": "training PPT assessment comparison visual asset",
+  "goal": "Create a 16:9 training-slide visual comparing static score evaluation with dynamic growth-state assessment. Final labels are added by PPT tooling.",
+  "source_page": {
+    "slide_id": "P12",
+    "teaching_goal": "Help the audience see that intelligent assessment identifies learning process and growth state, not only final score."
+  },
+  "layout": {
+    "page_type": "two-column contrast with transition arrow",
+    "left_column": {
+      "role": "traditional result scoring",
+      "visual": "cold static score sheet silhouette, ranking bars, checkmarks and red marks as abstract shapes, no readable numbers",
+      "mood": "flat, final, judgment-oriented"
+    },
+    "right_column": {
+      "role": "AI + SPGM growth-state assessment",
+      "visual": "dynamic growth map with student learning path, thinking nodes, teacher feedback loops, AI analysis lines, state changes over time",
+      "mood": "warm, developmental, process-oriented"
+    },
+    "transition": "arrow from result-only to process-plus-state",
+    "title_safe_area": "top area available for slide title and subtitle"
+  },
+  "style": {
+    "rendering": "clean educational diagram slide, human-centered technology",
+    "palette": "cool gray for left, warm blue-green and gold for right, off-white canvas",
+    "information_density": "medium-high with readable visual hierarchy"
+  },
+  "constraints": {
+    "must_keep": [
+      "left side feels static and score-based",
+      "right side shows learning path, teacher feedback, and AI analysis as connected system",
+      "contrast is understandable without final text",
+      "no generated paragraphs or fake score details"
+    ],
+    "avoid": [
+      "generic dashboard",
+      "unreadable tiny tables",
+      "teacher/student elements missing from the growth map",
+      "overly corporate analytics style",
+      "logos or watermarks"
+    ]
+  }
+}
+```
+
+## T-P18 Advisor: AI教育变革的逻辑链
+
+```json
+{
+  "type": "training PPT logic-chain visual asset",
+  "goal": "Create a 16:9 summary-slide visual showing the logic chain of AI education transformation. Final text labels are added outside the generated image.",
+  "source_page": {
+    "slide_id": "P18",
+    "teaching_goal": "Help the audience review the full lecture as one cognitive chain."
+  },
+  "layout": {
+    "page_type": "horizontal logic chain",
+    "node_count": 5,
+    "nodes": [
+      {
+        "id": 1,
+        "role": "AI ecosystem change",
+        "visual": "education ecosystem network icon: school, students, teachers, AI tools"
+      },
+      {
+        "id": 2,
+        "role": "expert cognitive model",
+        "visual": "structured thinking model / layered knowledge core"
+      },
+      {
+        "id": 3,
+        "role": "CCPE",
+        "visual": "expert thinking entering AI through structured protocol layers"
+      },
+      {
+        "id": 4,
+        "role": "SPGM and assessment innovation",
+        "visual": "matrix / growth-state map with teacher-student interaction"
+      },
+      {
+        "id": 5,
+        "role": "human-AI collaborative education ecology",
+        "visual": "school ecosystem with teacher, student, AI assistant, feedback loops"
+      }
+    ],
+    "connectors": "thin luminous arrows from left to right, each node visually distinct but same style",
+    "title_safe_area": "top band empty enough for title and core conclusion"
+  },
+  "style": {
+    "rendering": "elegant keynote presentation visual, educational technology, clean vector-like illustration",
+    "palette": "academic blue, warm gold accents, soft off-white background",
+    "information_density": "medium, designed for overlay text"
+  },
+  "constraints": {
+    "must_keep": [
+      "exactly five main nodes",
+      "left-to-right sequence is unambiguous",
+      "each node has a distinct visual role",
+      "chain ends in education ecology, not generic AI future"
+    ],
+    "avoid": [
+      "fake readable labels inside the image",
+      "generic AI marketing network",
+      "node count drifting from five",
+      "overly complex background",
+      "logos or watermarks"
+    ]
+  }
+}
+```
+
+## C-01 Advisor: 人物形象目录边界样本
+
+```json
+{
+  "type": "makeup-still directory boundary sample",
+  "goal": "Create a prompt that would belong under projects/<project-id>/makeup-still/ only when a project has a recurring human or character identity system. This is a low-weight boundary sample, not a core use case.",
+  "subject": {
+    "identity": "50-year-old East Asian male knowledge worker",
+    "fixed_markers": ["black rectangular glasses", "natural stubble", "tired but steady eyes", "slightly messy hair"],
+    "wardrobe_hint": "dark jacket or coat, understated"
+  },
+  "composition": {
+    "shot": "neutral head-and-shoulders identity reference",
+    "background": "plain gray or lightly blurred office-like neutral background",
+    "use": "identity_ref for future generated visuals, not final slide artwork"
+  },
+  "style": {
+    "rendering": "photorealistic reference portrait, natural light, no glamour"
+  },
+  "constraints": {
+    "must_keep": [
+      "stable identity markers",
+      "usable as a reference asset",
+      "plain enough for reuse"
+    ],
+    "avoid": [
+      "turning this into a universal workflow for non-character projects",
+      "hero poster or fashion portrait",
+      "readable text, logo, watermark"
+    ]
+  }
+}
+```
--- a/investigations/2026-06-23-gpt-image-2-advisor-prompt-quality-v2-broader-use-cases/outputs/direct-prompts.md
+++ b/investigations/2026-06-23-gpt-image-2-advisor-prompt-quality-v2-broader-use-cases/outputs/direct-prompts.md
@ -0,0 +1,35 @@
+# Direct Prompts
+
+These are the baseline direct prompts. For the science-video and training-PPT cases, they are taken from or kept close to the GPT case-file prompt style.
+
+## V-S01 Direct: 被浪拍下水的人
+
+A cinematic metaphorical illustration of ordinary people being swept into a huge dark blue ocean wave filled with subtle AI symbols, video thumbnails, chat windows, code fragments, and glowing digital particles. One person clutches a bright inflatable ring labeled conceptually as AI without readable text. Dramatic lighting, deep blue water, high tension, human-centered technology theme, editorial visual essay style, no readable text, 16:9.
+
+## V-S03 Direct: 游泳圈与太平洋
+
+A small human figure holding a bright inflatable ring symbolizing AI, standing at the edge of shallow water and facing a vast dark ocean under a stormy sky. The ocean feels enormous, dangerous, and overwhelming. Strong contrast between the playful floating ring and the terrifying deep sea. Cinematic metaphorical illustration, dark blue palette, dramatic clouds, no readable text, 16:9.
+
+## V-S05 Direct: 逻辑沙堡
+
+A highly detailed sandcastle on a beach, shaped like a complex business plan with towers, bridges, charts, and fragile pathways, while a huge ocean wave is about to crash into it. Scattered elegant business plan pages and diagrams lie nearby, but no readable text. Metaphorical cinematic illustration, dramatic tension, golden beach light with dark blue wave, no readable text, 16:9.
+
+## V-S07 Direct: 防忽悠三板斧
+
+A metaphorical scene of a small vehicle about to rush toward a dangerous cliff and deep ocean, stopped by three glowing brake gates. The three gates are represented by icons: a calculator for risk calculation, a cyber watchdog for red-team attack, and a signed cold contract for stop-loss rules. Cinematic educational illustration, dark blue and amber palette, no readable text, 16:9.
+
+## T-P8 Direct: 提示词工程已死？认知工程才刚开始
+
+Split-screen conceptual illustration: on the left, scattered old prompt cards and broken magic scrolls fading away; on the right, a glowing structured cognitive core shaped like a brain with connected process loops, evaluation nodes, and domain knowledge layers, futuristic academic style, clean composition, no readable text, 16:9.
+
+## T-P12 Direct: 从分数评价到成长状态评价
+
+Split-screen educational assessment illustration: left side a traditional static score sheet and ranking chart, right side a dynamic growth map with student learning paths, thinking nodes, teacher feedback loops, and AI analysis lines, warm human-centered technology style, no readable text, 16:9.
+
+## T-P18 Direct: AI教育变革的逻辑链
+
+A clean horizontal logic chain diagram for educational AI transformation, five connected glowing nodes representing AI ecosystem, cognitive engineering, educational model, assessment innovation, and human-AI collaborative school, elegant keynote presentation style, no readable text, 16:9.
+
+## C-01 Direct: 人物形象目录边界样本
+
+Create a realistic reference portrait of a recurring video protagonist: a 50-year-old East Asian male knowledge worker with black rectangular glasses, natural stubble, tired but steady eyes, slightly messy hair, and a dark jacket. Clean neutral background, realistic lighting, no readable text.
--- a/investigations/2026-06-23-gpt-image-2-advisor-prompt-quality-v2-broader-use-cases/提示词质量综合调研报告.md
+++ b/investigations/2026-06-23-gpt-image-2-advisor-prompt-quality-v2-broader-use-cases/提示词质量综合调研报告.md
@ -0,0 +1,232 @@
+# 提示词质量综合调研报告
+
+## 结论
+
+在只评估提示词质量、不生成图片的前提下，`gpt-image-2` Advisor 工作流仍然明显优于直接 prompt。
+
+这次结论比 v1 更稳，因为样本权重已调整为：
+
+```text
+典型视频分镜(科普) 55%
+典型培训AI PPT     35%
+人物造型           10%
+```
+
+并且科普视频和培训 PPT 的 baseline 不是弱 prompt，而是直接采用或贴近 GPT 知识库两个典型案例中的已有生图 prompt。
+
+综合加权结果：
+
+| Prompt 类型 | 加权平均分 / 30 |
+| --- | ---: |
+| 直接 prompt | 21.6 |
+| Advisor prompt | 27.9 |
+
+核心判断：
+
+> Advisor 的优势不主要在“更会写画面形容词”，而在“把视频 / PPT 生产所需的结构、约束、版面安全区、文字责任边界和失败风险写进 prompt”。
+
+## 调研依据
+
+本报告使用两个指定案例：
+
+- `30_CASE_典型视频分镜案例.md`
+- `31_CASE_典型培训AI_PPT案例.md`
+
+关键案例规则包括：
+
+- 科普视频图片不负责正文文字，页面文字由 PPT / 视频页面层渲染。
+- 每个科普分镜只承载一个核心逻辑点，画面必须承担隐喻表达。
+- 培训 PPT 页不是镜头，而是教学单元，需要支持讲师停留、解释、提问和迁移。
+- 培训 PPT 的页面设计、讲师备注和生图 prompt 必须分开。
+
+这些规则与 `gpt-image-2` Advisor 方法匹配：将图像请求拆成 `type / goal / layout / subject / scene / style / constraints`，并避免让图像模型生成最终可读正文。
+
+## 评分 Rubric
+
+每个 case 满分 30 分，6 个维度各 0-5 分。
+
+| 维度 | 说明 |
+| --- | --- |
+| R1 媒介适配 | 是否知道这是科普视频、培训 PPT、还是人物参考，而不是泛图。 |
+| R2 结构控制 | 是否明确布局、景别、前中后景、对比栏、节点链路或标题安全区。 |
+| R3 文字责任边界 | 是否避免让图像模型生成正文、标签、分数、中文说明。 |
+| R4 隐喻 / 教学逻辑 | 是否服务于单一核心逻辑点或教学目标。 |
+| R5 失败预防 | 是否明确常见失败：过度装饰、假文字、图标不清、隐喻漂移。 |
+| R6 复用与迭代 | 是否适合落入 video-workbench 的版本化 prompt / review 流程。 |
+
+## 分组结果
+
+### A. 科普视频分镜，权重 55%
+
+| Case | Direct / 30 | Advisor / 30 | 主要差异 |
+| --- | ---: | ---: | --- |
+| V-S01 被浪拍下水的人 | 23 | 28 | Advisor 明确 opening hook、标题安全区、AI 元素只能用不可读符号。 |
+| V-S03 游泳圈与太平洋 | 22 | 28 | Advisor 强化人物小、海洋占比、浅水/深水边界，避免旅游海报化。 |
+| V-S05 逻辑沙堡 | 23 | 28 | Advisor 把“精美但脆弱”和“浪头迫近”拆成可检查构图约束。 |
+| V-S07 防忽悠三板斧 | 22 | 27 | Advisor 明确三道 gate 的数量、图标角色和车辆-闸门-风险的运动方向。 |
+
+组平均：
+
+| Prompt 类型 | 平均分 / 30 |
+| --- | ---: |
+| Direct | 22.5 |
+| Advisor | 27.8 |
+
+科普视频里的直接 prompt 已经不错，特别是隐喻清楚、画面感强。但 Advisor 在三个地方更稳：
+
+- 更适合后续叠加页面文字，因为会主动保留 title-safe area。
+- 更少依赖“形容词堆叠”，更多使用可检查结构。
+- 更适合 review，因为 acceptance criteria 可以直接从 constraints 中抽取。
+
+### B. 培训 AI PPT，权重 35%
+
+| Case | Direct / 30 | Advisor / 30 | 主要差异 |
+| --- | ---: | ---: | --- |
+| T-P8 提示词工程 vs 认知工程 | 21 | 28 | Advisor 把左右对比、转化桥、页面文字责任拆开，教学用途更清楚。 |
+| T-P12 分数评价 vs 成长状态评价 | 21 | 28 | Advisor 明确静态评分与动态成长地图的视觉差异，避免泛数据看板。 |
+| T-P18 AI教育变革逻辑链 | 20 | 29 | Advisor 强制五节点链路、节点角色和顺序，避免泛 AI 营销图。 |
+
+组平均：
+
+| Prompt 类型 | 平均分 / 30 |
+| --- | ---: |
+| Direct | 20.7 |
+| Advisor | 28.3 |
+
+培训 PPT 是 Advisor 提升最大的场景。原因很具体：
+
+- 培训页经常有“页面结构 + 可编辑文字 + 视觉资产”三层，直接 prompt 容易混在一起。
+- Advisor 会把 `layout`、`title_safe_area`、`no final readable text`、`post-use` 关系写清楚。
+- 对模型页、对比页、逻辑链页，Advisor 的字段化结构更像可复用的页面资产规范。
+
+### C. 人物造型，权重 10%
+
+| Case | Direct / 30 | Advisor / 30 | 主要差异 |
+| --- | ---: | ---: | --- |
+| C-01 人物形象目录边界样本 | 20 | 27 | Advisor 明确这是 `makeup-still/` 的低权重边界样本，不扩张成通用流程。 |
+
+人物造型不应再主导本次改造判断。它的价值只在于：当项目确实有人物或角色连续性时，Advisor 能帮助定义 `identity_ref`、目录边界和复用目的。
+
+## 加权计算
+
+| Group | Weight | Direct Avg | Advisor Avg | Direct Weighted | Advisor Weighted |
+| --- | ---: | ---: | ---: | ---: | ---: |
+| 科普视频分镜 | 55% | 22.5 | 27.8 | 12.4 | 15.3 |
+| 培训 AI PPT | 35% | 20.7 | 28.3 | 7.2 | 9.9 |
+| 人物造型 | 10% | 20.0 | 27.0 | 2.0 | 2.7 |
+| **Total** | **100%** |  |  | **21.6** | **27.9** |
+
+## 质量差异归因
+
+### 1. Advisor 更能保住“图片不负责正文文字”
+
+两个指定案例都强调：页面正文、标题、金句由 PPT / 视频页面层渲染，图片 prompt 只生成背景、视觉隐喻或图示资产。
+
+直接 prompt 虽然常写 `no readable text`，但经常同时要求“score sheet / ranking chart / logic chain / prompt cards”等文本性对象，模型容易生成假字。
+
+Advisor prompt 会进一步说明：
+
+- 只生成 score-sheet silhouette，不生成具体分数。
+- 只生成 abstract prompt cards，不生成可读 prompt。
+- 只生成 node visuals，不生成节点文字。
+- 保留 title-safe area 给后续页面层。
+
+这对培训 PPT 和科普视频都很关键。
+
+### 2. Advisor 更适合“教学单元”
+
+培训 PPT 不是单纯好看，而是要让讲师能停下来讲。
+
+直接 prompt 多数只给一张概念图。Advisor 会补足：
+
+- `teaching_goal`
+- `page_type`
+- `left_column / right_column`
+- `transition`
+- `title_safe_area`
+- `constraints`
+
+这些字段让图像资产更像“页面组件”，不是独立插画。
+
+### 3. Advisor 更适合“分镜资产”
+
+科普视频分镜需要画面推进和隐喻表达。直接 prompt 的画面描述强，但不总是说明：
+
+- 这个镜头在视频里的角色是什么；
+- 需要给叠字留哪里；
+- 哪些符号只能作为不可读元素；
+- 怎么避免隐喻漂移；
+- 后续 review 如何判断通过。
+
+Advisor prompt 会把这些都显式化。
+
+### 4. Advisor 的成本是更长、更工程化
+
+Advisor prompt 不适合所有场景。对于很简单的单主体图，直接 prompt 足够。
+
+但在 video-workbench 高频任务中，主要是科普视频和培训 PPT。这些任务不是“一张漂亮图”，而是“可叠字、可讲解、可复用、可 review 的生产资产”。在这个条件下，Advisor 的冗长是成本，也是控制力来源。
+
+## 对改造前提的判断
+
+原始前提是：
+
+> Codex 具有 `GPT image 2 skill`，其 Advisor 模式下，如果再叠加 GPT 的知识库，出的图片提示词质量应该比 GPT 高。
+
+本次调研的判断：
+
+```text
+在提示词质量层面：成立。
+在最终图像质量层面：尚未验证。
+```
+
+更准确的表述应改为：
+
+> GPT 负责输出导演卡、教学页设计、分镜意图和验收标准；Codex 使用 `gpt-image-2` Advisor 将这些内容转成结构化、可复用、可 review 的图片 prompt。对科普视频和培训 PPT，这类 prompt 的生产质量明显高于直接生成一段生图 prompt。
+
+## 对目录和知识库的影响
+
+### 1. `investigations/` 是合适的
+
+这类内容是前提验证，不是项目资产，也不是稳定知识库。放在：
+
+```text
+investigations/<date>-<topic>/
+```
+
+比放在 `tmp/` 更可追踪，比直接放 `docs/` 更谨慎。
+
+### 2. 稳定知识应进入 `docs/`
+
+如果接受本报告结论，再把稳定规则整理进 `video-workbench/docs/`，例如：
+
+```text
+docs/image-prompt-advisor-workflow.md
+docs/visual-asset-prompt-rubric.md
+docs/page-text-vs-image-prompt-boundary.md
+docs/makeup-still-directory-policy.md
+```
+
+其中前 3 个比 `makeup-still` 更重要，因为科普和培训才是主场景。
+
+### 3. `makeup-still/` 应是条件目录
+
+建议保留你的设想，但不要作为所有项目默认目录。
+
+```text
+projects/<project-id>/makeup-still/
+```
+
+只在人、角色、讲师形象、IP 形象需要跨镜头 / 跨页保持一致时出现。
+
+## 后续验证建议
+
+当前只关心提示词质量，因此本报告可以作为进入“分工与目录改造草案”的前置依据。
+
+如果后面要验证最终图像质量，再做 Level 2：
+
+1. 从本报告选 2 个科普视频 prompt、2 个培训 PPT prompt。
+2. 同平台、同参数分别出图。
+3. 用同一 acceptance criteria 评估成图。
+4. 记录 Advisor 是否实际降低假文字、构图漂移和隐喻失焦。
+
+本轮不需要做 Level 2。
--- a/investigations/2026-06-23-gpt-image-2-advisor-prompt-quality/README.md
+++ b/investigations/2026-06-23-gpt-image-2-advisor-prompt-quality/README.md
@ -0,0 +1,51 @@
+# GPT Image 2 Advisor Prompt Quality Investigation
+
+## Purpose
+
+This investigation tests a workflow premise before any GPT / video-workbench restructuring:
+
+> If Codex uses the `gpt-image-2` skill in Advisor mode and combines it with the accepted GPT director knowledge, the resulting image prompts should be stronger than ordinary GPT direct prompts.
+
+This is a prompt-text quality test only. It does not validate final image quality because no image generation is performed.
+
+## Scope
+
+The test compares two prompt styles across five representative video-workbench scenes:
+
+1. Direct prompt: a normal prompt written without consulting `gpt-image-2` templates.
+2. Advisor prompt: a prompt rendered after reading the relevant `gpt-image-2` skill guidance and templates.
+
+The current environment was checked with the skill's mode detector. It returned `B-or-C`, with Garden disabled and no local API key. This investigation forces Mode C / Advisor: save and compare prompts only.
+
+## Skill References Read
+
+- `gpt-image-2/SKILL.md`
+- `references/prompt-writing.md`
+- `references/portraits-and-characters/professional-portrait.md`
+- `references/portraits-and-characters/character-sheet.md`
+- `references/portraits-and-characters/pose-reference-sheet.md`
+- `references/storyboards-and-sequences/cinematic-storyboard-grid.md`
+- `references/scenes-and-illustrations/concept-scene.md`
+- `references/slides-and-visual-docs/dense-explainer-slides.md`
+
+## Files
+
+- `mode-check.json`: local mode detection result.
+- `cases.md`: test cases and expected prompt risks.
+- `outputs/direct-prompts.md`: baseline direct prompts.
+- `outputs/advisor-prompts.md`: Advisor-style prompts.
+- `comparison.md`: rubric, scores, and conclusion.
+
+## Directory Decision For Future Work
+
+This investigation uses a root-level `investigations/` directory because it is validating an assumption, not running a production project or automated test suite.
+
+If this premise is accepted, stable workflow documentation should move into `docs/`. Project-specific outputs should remain under `projects/<project-id>/`.
+
+For projects with recurring human character imagery, this investigation supports adding:
+
+```text
+projects/<project-id>/makeup-still/
+```
+
+Use it only when a project has a character / human identity package. It should hold makeup-still prompts, generated or externally produced makeup-still images, reference-role notes, and acceptance records. Projects without visible characters do not need this folder.
--- a/investigations/2026-06-23-gpt-image-2-advisor-prompt-quality/cases.md
+++ b/investigations/2026-06-23-gpt-image-2-advisor-prompt-quality/cases.md
@ -0,0 +1,85 @@
+# Test Cases
+
+## Evaluation Question
+
+Does the Advisor workflow produce prompts that are more controllable, reusable, and aligned with video-workbench production needs than direct prompting?
+
+## Case T01: Character Identity Makeup Still
+
+Goal: create a stable identity reference image for a recurring MV protagonist.
+
+Input:
+
+- 50-year-old East Asian male traveler.
+- Black rectangular glasses, stubble, tired but steady eyes, slightly messy hair.
+- Dark long coat, old backpack strap visible.
+- Not heroic, not fashion, not influencer portrait.
+- Output should become `identity_ref` for later shots.
+
+Main risk:
+
+- If the prompt over-cinematizes the result, the identity anchor becomes too stylized and hard to reuse.
+- If the prompt under-specifies face and texture, later shots drift.
+
+## Case T02: Wardrobe / Body Fullbody Anchor
+
+Goal: create a wardrobe and body silhouette reference, not a face reference.
+
+Input:
+
+- Same traveler system as T01.
+- Full body or nearly full body.
+- Long dark coat length, old backpack, trouser cuffs, muddy boots.
+- Face may be less important; outfit continuity is the goal.
+
+Main risk:
+
+- Prompt tries to solve face, wardrobe, mood, and scene at once.
+- The result becomes another portrait instead of a silhouette / costume reference.
+
+## Case T03: MV Side-Walking Road Shot
+
+Goal: make a production still for an MV shot where the character walks along a cold road.
+
+Input:
+
+- Use identity_ref for face system, wardrobe_ref for outfit, pose_ref for side / three-quarter walking.
+- Environment occupies at least 60%.
+- The character should not look into camera.
+- Vertical 9:16, road-movie mood, restrained fatigue.
+
+Main risk:
+
+- The identity reference pulls the result back into centered portrait.
+- The road scene loses character consistency.
+
+## Case T04: Low-Angle Muddy Boots Detail
+
+Goal: make a close detail shot where identity is carried by boots, coat hem, mud, and bodily cost.
+
+Input:
+
+- Low angle near the ground.
+- Muddy boots stepping through wet road.
+- Coat hem and backpack strap may appear, face is not required.
+- Vertical 9:16, tactile road-movie detail.
+
+Main risk:
+
+- Overusing the face identity reference harms the shot.
+- The image becomes a generic boot photo without narrative continuity.
+
+## Case T05: High-Density Explainer Slide Visual
+
+Goal: test a non-character visual common in video-workbench: one dense page explaining a concept.
+
+Input:
+
+- Topic: why AI image production needs staged visual anchors.
+- Style: clean training slide / explainer page.
+- Output image should be a visual background or slide asset; readable text should be handled outside the generated image.
+
+Main risk:
+
+- Direct prompt asks the image model to render too much text.
+- The result may be visually attractive but structurally useless for a slide.
--- a/investigations/2026-06-23-gpt-image-2-advisor-prompt-quality/comparison.md
+++ b/investigations/2026-06-23-gpt-image-2-advisor-prompt-quality/comparison.md
@ -0,0 +1,86 @@
+# Comparison
+
+## Rubric
+
+Scores are subjective prompt-text scores from 0 to 5. They judge whether the prompt is likely to be controllable and reusable in video-workbench. They do not judge generated image quality.
+
+| Dimension | Meaning |
+| --- | --- |
+| D1 Intent Decomposition | Separates goal, subject, scene, and output use. |
+| D2 Reference Role Control | Clarifies which reference preserves identity, wardrobe, pose, or style. |
+| D3 Composition Specificity | Defines camera, framing, aspect ratio, environment ratio, and motion. |
+| D4 Failure Prevention | Names predictable failure modes and avoids them. |
+| D5 Workflow Fit | Fits `projects/<project-id>/...` style production records and later review. |
+| D6 Model Usability | Gives image models concrete visual instructions without overloading readable text. |
+
+## Scores
+
+| Case | Direct Total / 30 | Advisor Total / 30 | Main Difference |
+| --- | ---: | ---: | --- |
+| T01 Identity makeup still | 20 | 27 | Advisor treats the result as `identity_ref`, not a final cinematic scene. |
+| T02 Wardrobe fullbody anchor | 18 | 27 | Advisor explicitly downgrades face likeness and prioritizes outfit geometry. |
+| T03 MV side-walking shot | 19 | 29 | Advisor separates identity, wardrobe, pose, and style references and guards against portrait drift. |
+| T04 Muddy boots detail | 22 | 28 | Advisor correctly avoids using identity as the primary reference. |
+| T05 Dense explainer slide | 18 | 27 | Advisor prevents the common error of asking the image model to render final readable text. |
+
+Average:
+
+| Prompt Type | Average / 30 |
+| --- | ---: |
+| Direct | 19.4 |
+| Advisor | 27.6 |
+
+## Findings
+
+At the prompt-text level, the Advisor workflow is materially stronger than direct prompting for this use case.
+
+The main improvement is not prettier language. It is production control:
+
+- It states whether an output is an identity anchor, wardrobe anchor, final shot, detail shot, or slide background.
+- It splits references into roles instead of treating all references as the same kind of input.
+- It encodes known failures from the prior run, especially portrait drift and overusing face identity.
+- It creates prompts that can be versioned, reviewed, and revised per shot.
+
+This supports the restructuring premise:
+
+```text
+GPT should output director cards and acceptance standards.
+Codex / video-workbench should render final image prompts using gpt-image-2 Advisor logic.
+```
+
+## Limits
+
+This does not prove final image quality. It proves only that the Advisor prompt artifacts are higher-quality production inputs under a text rubric.
+
+The next validation level should be:
+
+1. Pick 2-3 prompts from this investigation.
+2. Generate images on the same image platform.
+3. Judge image outcomes with the same acceptance criteria.
+4. Record whether the Advisor prompts actually reduce failure rates.
+
+## Path Implications
+
+The investigation supports adding a project-local character package folder for human-character projects:
+
+```text
+projects/<project-id>/makeup-still/
+  identity-prompt-v1.md
+  identity-ref-v1.png
+  wardrobe-prompt-v1.md
+  wardrobe-ref-v1.png
+  pose-ref-notes.md
+  acceptance-notes.md
+```
+
+This folder should exist only when a project needs a recurring human / character identity system. Non-character projects should not carry the folder.
+
+Stable Codex-side knowledge should not be buried in a single project. After owner acceptance, summarize this into `docs/` as video-workbench operating knowledge, likely:
+
+```text
+docs/image-prompt-advisor-workflow.md
+docs/makeup-still-directory-policy.md
+docs/reference-strategy-and-review-rubric.md
+```
+
+Do not write those stable docs before the owner accepts this investigation result.
--- a/investigations/2026-06-23-gpt-image-2-advisor-prompt-quality/mode-check.json
+++ b/investigations/2026-06-23-gpt-image-2-advisor-prompt-quality/mode-check.json
@ -0,0 +1,12 @@
+{
+  "checked_at": "2026-06-23",
+  "command": "node C:\\Users\\wangq\\.agents\\skills\\gpt-image-2\\scripts\\check-mode.js --json",
+  "mode": "B-or-C",
+  "recommendation": "host-or-advisor",
+  "garden_mode_enabled": false,
+  "has_api_key": false,
+  "base_url": "https://api.openai.com/v1",
+  "model": "gpt-image-2",
+  "env_flag_value": "(unset)",
+  "test_decision": "Force Mode C / Advisor for this investigation. Do not generate images."
+}
--- a/investigations/2026-06-23-gpt-image-2-advisor-prompt-quality/outputs/advisor-prompts.md
+++ b/investigations/2026-06-23-gpt-image-2-advisor-prompt-quality/outputs/advisor-prompts.md
@ -0,0 +1,283 @@
+# Advisor Prompts
+
+These prompts apply `gpt-image-2` Advisor guidance: choose the nearest template family, split fields, define must-keep / avoid constraints, and make reference roles explicit.
+
+## T01 Advisor: Character Identity Makeup Still
+
+```json
+{
+  "type": "character identity makeup still / reusable identity reference",
+  "goal": "Generate a stable identity reference image for a recurring MV protagonist. The image is not the final MV scene; it will be used as identity_ref in later image-to-image shots.",
+  "subject": {
+    "identity": "50-year-old East Asian male traveler",
+    "face": {
+      "glasses": "black rectangular glasses",
+      "hair": "short dark hair, slightly messy, natural not styled",
+      "facial_hair": "light stubble, uneven and realistic",
+      "skin": "natural skin texture, visible pores, no plastic smoothing",
+      "eyes": "tired but steady, restrained emotional weight"
+    },
+    "expression": {
+      "mood": "quiet resilience, fatigue contained rather than dramatized",
+      "gaze": "neutral forward gaze, not heroic, not influencer-like"
+    },
+    "wardrobe_visible": "dark long coat collar and upper chest, old backpack strap visible but secondary"
+  },
+  "composition": {
+    "shot": "head-and-shoulders or upper torso portrait",
+    "framing": "centered enough for identity reuse, simple background, no dramatic scene action",
+    "aspect_ratio": "3:4 or vertical 9:16",
+    "background": "plain cool gray outdoor overcast background, lightly blurred"
+  },
+  "style": {
+    "image_type": "photorealistic makeup still / character reference photo",
+    "lighting": "soft overcast natural light, low contrast, no glamour rim light",
+    "color_grading": "cool gray road-movie palette, muted skin tones",
+    "camera": "85mm portrait lens feel, realistic depth of field"
+  },
+  "constraints": {
+    "must_keep": [
+      "face identity must be stable and reusable",
+      "black rectangular glasses are a fixed identity marker",
+      "stubble, age, and tired eyes must remain visible",
+      "wardrobe should hint at the traveler system without becoming a fashion shoot"
+    ],
+    "avoid": [
+      "hero poster",
+      "fashion editorial portrait",
+      "smiling influencer headshot",
+      "excessive wrinkles or caricatured aging",
+      "dramatic fantasy lighting",
+      "readable text, logo, watermark"
+    ]
+  },
+  "usage_note": "Use this output as identity_ref only. Do not use it as the sole reference for wide walking shots because it may pull later generations toward portrait framing."
+}
+```
+
+## T02 Advisor: Wardrobe / Body Fullbody Anchor
+
+```json
+{
+  "type": "wardrobe and body silhouette reference",
+  "goal": "Generate a reusable wardrobe_ref / body_ref for later MV shots. Face similarity is secondary; clothing geometry, body proportions, and traveler silhouette are primary.",
+  "subject": {
+    "character_system": "same 50-year-old East Asian male traveler system as identity_ref",
+    "visibility": "full body, head to boots, standing in neutral pose",
+    "priority": {
+      "highest": "coat length, backpack shape, trouser cuffs, muddy boots, body proportion",
+      "medium": "age impression, glasses presence",
+      "lower": "precise facial likeness"
+    },
+    "wardrobe": {
+      "coat": "dark long coat reaching around knee level, worn fabric, not stylish new fashion",
+      "backpack": "old practical backpack with visible shoulder straps, slightly sagging",
+      "pants": "dark worn trousers, relaxed fit, cuffs near boots",
+      "boots": "sturdy muddy travel boots, scuffed leather or canvas"
+    }
+  },
+  "composition": {
+    "shot": "full-body neutral reference photo",
+    "camera": "straight-on eye-level, consistent scale, no dramatic perspective",
+    "pose": "standing or slight walking-ready stance, arms natural, body readable",
+    "background": "plain light gray or muted outdoor wall, minimal distractions",
+    "aspect_ratio": "3:4 vertical"
+  },
+  "style": {
+    "image_type": "photorealistic costume continuity reference",
+    "lighting": "even soft daylight",
+    "rendering": "clear clothing edges and material texture, not painterly"
+  },
+  "constraints": {
+    "must_keep": [
+      "full body visible including boots",
+      "coat length and backpack silhouette clearly readable",
+      "outfit looks used, practical, and road-worn",
+      "single subject only"
+    ],
+    "avoid": [
+      "tight portrait crop",
+      "runway fashion styling",
+      "clean luxury clothes",
+      "strong action pose that hides outfit structure",
+      "extra bags or props that confuse continuity",
+      "readable text, logo, watermark"
+    ]
+  },
+  "usage_note": "Use this output as wardrobe_ref. Pair it with identity_ref for face-bearing shots and pose_ref for action/composition shots."
+}
+```
+
+## T03 Advisor: MV Side-Walking Road Shot
+
+```json
+{
+  "type": "single cinematic MV storyboard still",
+  "goal": "Generate one vertical 9:16 MV production still from a director card. The shot should feel like a road-movie scene, not a character portrait.",
+  "reference_strategy": {
+    "identity_ref": "use only to preserve the recurring character system: 50-year-old East Asian male, black rectangular glasses, stubble, tired steady face",
+    "wardrobe_ref": "use to preserve long dark coat, old backpack, trouser cuffs, muddy boots, body silhouette",
+    "pose_ref": "use for side or three-quarter walking posture, gaze toward the road, no direct eye contact",
+    "style_ref": "cold morning road-movie still, muted gray-blue light, restrained fatigue"
+  },
+  "scene": {
+    "time": "early morning, cold air, thin mist",
+    "location": "empty rural or suburban road leading into distance",
+    "weather": "damp road surface, light frost or cold haze",
+    "emotional_function": "the character is still moving after hardship; fatigue becomes forward motion"
+  },
+  "composition": {
+    "aspect_ratio": "vertical 9:16",
+    "shot_size": "medium-wide to wide shot, full body or knee-up",
+    "camera_angle": "eye-level or slightly low, not portrait lens close-up",
+    "character_position": "off-center on lower third or side third",
+    "environment_ratio": "at least 60% of the image is road, mist, sky, and surrounding space",
+    "directionality": "road line leads away from camera, character walking along it",
+    "gaze": "looking toward the road ahead, no direct camera gaze"
+  },
+  "style": {
+    "image_type": "photorealistic cinematic still",
+    "lens": "35mm or 50mm road-movie framing, natural perspective",
+    "lighting": "soft cold dawn light, low contrast, realistic shadows",
+    "color_grading": "muted blue-gray with slight warm skin undertone"
+  },
+  "constraints": {
+    "must_keep": [
+      "same character system as identity_ref",
+      "same wardrobe system as wardrobe_ref",
+      "body in motion, not posing",
+      "environment dominates the frame",
+      "road depth and walking direction are clear"
+    ],
+    "avoid": [
+      "centered front-facing portrait",
+      "direct eye contact",
+      "heroic poster stance",
+      "fashion photo",
+      "clean new clothing",
+      "overly dramatic rainstorm",
+      "readable text, logo, watermark"
+    ]
+  },
+  "failure_guard": "If the model over-prioritizes face identity and produces a portrait, reduce identity_ref weight and strengthen pose_ref + environment_ratio in the next prompt."
+}
+```
+
+## T04 Advisor: Low-Angle Muddy Boots Detail
+
+```json
+{
+  "type": "cinematic MV detail shot / tactile continuity shot",
+  "goal": "Generate a vertical detail still where the character's struggle is carried by boots, mud, coat hem, and movement rather than face identity.",
+  "reference_strategy": {
+    "primary_ref": "wardrobe_ref or boots_ref",
+    "secondary_ref": "pose/action reference for stepping through wet mud",
+    "identity_ref": "not needed as primary; use only if the image tool supports low-weight reference roles"
+  },
+  "subject": {
+    "main": "muddy travel boots stepping into wet mud on a cold road",
+    "supporting_details": [
+      "dark long coat hem partially visible",
+      "old backpack strap or swinging shadow may appear at frame edge",
+      "mud splashes and wet gravel carry the sense of effort"
+    ],
+    "face_visibility": "no face required"
+  },
+  "composition": {
+    "aspect_ratio": "vertical 9:16",
+    "camera": "very low angle near ground, close to boots",
+    "shot_size": "extreme low close-up with enough surrounding road texture",
+    "motion": "one boot compressing mud, the other lifting forward",
+    "depth": "road surface recedes into soft misty background"
+  },
+  "style": {
+    "image_type": "photorealistic cinematic road-movie detail",
+    "lighting": "cold natural dawn light, wet surface reflections",
+    "texture": "mud, water, scuffed boot leather, coat fabric edge",
+    "color_grading": "muted earth brown, wet black, cold blue-gray highlights"
+  },
+  "constraints": {
+    "must_keep": [
+      "boots and mud are the emotional subject",
+      "shot is not a portrait",
+      "same wardrobe language as the traveler system",
+      "tactile realism: mud has weight, water has reflections"
+    ],
+    "avoid": [
+      "showing a face",
+      "clean boots",
+      "generic fashion footwear ad",
+      "fantasy heroic lighting",
+      "excessive splash that feels staged",
+      "readable text, logo, watermark"
+    ]
+  },
+  "usage_note": "This shot should use wardrobe_ref / boots_ref as the main anchor. Overusing identity_ref is a known failure path."
+}
+```
+
+## T05 Advisor: High-Density Explainer Slide Visual
+
+```json
+{
+  "type": "high-density explainer slide background / visual system",
+  "goal": "Generate a clean training-slide visual explaining staged visual anchors in AI image production. The generated image should provide structure and icons; final readable text should be added by slide tooling, not by the image model.",
+  "topic": "why AI image production needs staged visual anchors",
+  "style": {
+    "format": "clean explainer slide / ponchi-e inspired diagram",
+    "blend": "training slide clarity + soft operational illustration",
+    "color_palette": "warm off-white background, deep gray lines, muted red accent, small blue-gray highlights",
+    "information_density": "medium-high but readable"
+  },
+  "layout": {
+    "aspect_ratio": "16:9",
+    "title_safe_area": "top band left mostly empty for real slide title added later",
+    "centerpiece": "a horizontal workflow spine with five visual stations",
+    "sections": [
+      {
+        "id": "1",
+        "visual_role": "identity anchor",
+        "image": "small portrait-card icon with face silhouette and glasses marker, no readable text"
+      },
+      {
+        "id": "2",
+        "visual_role": "wardrobe anchor",
+        "image": "full-body outfit card icon with coat, backpack, boots"
+      },
+      {
+        "id": "3",
+        "visual_role": "pose / composition anchor",
+        "image": "side-walking pose frame with road perspective"
+      },
+      {
+        "id": "4",
+        "visual_role": "prompt rendering",
+        "image": "structured prompt card with abstract lines, not readable text"
+      },
+      {
+        "id": "5",
+        "visual_role": "review and rerun",
+        "image": "comparison/checklist icon and circular arrow"
+      }
+    ],
+    "connector_style": "thin arrows between sections, light grid alignment, generous margins"
+  },
+  "constraints": {
+    "must_keep": [
+      "clear left-to-right process structure",
+      "visual hierarchy supports later overlay text",
+      "no dense generated text",
+      "icons are consistent in line weight and color",
+      "the diagram implies staged anchors before final image generation"
+    ],
+    "avoid": [
+      "tiny unreadable paragraphs",
+      "random AI robot imagery",
+      "too many colors",
+      "crowded margins",
+      "fake UI text, logo, watermark"
+    ]
+  },
+  "post_use_note": "After image generation, add real Chinese labels in slide/page tooling. Do not rely on the image model for final readable text."
+}
+```
--- a/investigations/2026-06-23-gpt-image-2-advisor-prompt-quality/outputs/direct-prompts.md
+++ b/investigations/2026-06-23-gpt-image-2-advisor-prompt-quality/outputs/direct-prompts.md
@ -0,0 +1,23 @@
+# Direct Prompts
+
+These prompts represent a competent direct GPT-style baseline without using `gpt-image-2` templates.
+
+## T01 Direct: Character Identity Makeup Still
+
+Create a realistic cinematic portrait of a 50-year-old East Asian male traveler. He has black rectangular glasses, short messy dark hair, light stubble, tired but steady eyes, and a quiet resilient expression. He wears a dark long coat with an old backpack strap visible. The mood is restrained, road-movie style, natural skin texture, soft overcast light, neutral background, no text, no logo, vertical portrait.
+
+## T02 Direct: Wardrobe / Body Fullbody Anchor
+
+Create a full-body reference image of the same 50-year-old East Asian male traveler. He wears a dark long coat, old backpack, worn trousers, and muddy boots. Show his body shape and outfit clearly from head to toe. The scene should be simple and neutral so the clothing can be reused as a reference. Realistic cinematic look, no text, no logo.
+
+## T03 Direct: MV Side-Walking Road Shot
+
+Using the traveler reference, create a vertical 9:16 cinematic MV still of him walking along a cold road in early morning. He is tired but still moving forward, wearing the same dark long coat, backpack, glasses, and muddy boots. The camera sees him from the side or three-quarter angle, not a front portrait. The environment should be wide and misty, with the road leading into the distance. No readable text.
+
+## T04 Direct: Low-Angle Muddy Boots Detail
+
+Create a vertical 9:16 cinematic close-up of the traveler's muddy boots stepping through wet mud on a road. Low angle near the ground, coat hem visible, rainwater and mud splashing, tired road-movie feeling. It should feel like the same traveler but without showing his face. Realistic, tactile, no text, no logo.
+
+## T05 Direct: High-Density Explainer Slide Visual
+
+Create a clean high-density explainer slide image about why AI image production needs staged visual anchors. Show a process from character identity to wardrobe, pose, scene, prompt, review, and rerun. Use simple icons and a clear flow, warm neutral colors, professional training style, no tiny unreadable text.
--- a/investigations/2026-06-23-text-to-image-image-to-image-execution-boundary/README.md
+++ b/investigations/2026-06-23-text-to-image-image-to-image-execution-boundary/README.md
@ -0,0 +1,13 @@
+# Text-To-Image / Image-To-Image Execution Boundary Investigation
+
+## Purpose
+
+This investigation records how Video Workbench should distinguish text-to-image and image-to-image work after GPT V2 stops producing final image prompts and Codex becomes the local prompt owner.
+
+## Scope
+
+This is a workflow and responsibility investigation. It does not generate images and does not claim final image quality.
+
+## Files
+
+- `文生图与图生图执行边界调研.md`: local workflow findings and recommended rules.
--- a/investigations/2026-06-23-text-to-image-image-to-image-execution-boundary/文生图与图生图执行边界调研.md
+++ b/investigations/2026-06-23-text-to-image-image-to-image-execution-boundary/文生图与图生图执行边界调研.md
@ -0,0 +1,226 @@
+# 文生图与图生图执行边界调研
+
+```yaml
+status: investigation
+date: 2026-06-23
+scope:
+  - text-to-image
+  - image-to-image
+  - video-workbench local execution workflow
+non_goals:
+  - 不直接生图
+  - 不验证最终图片质量
+  - 不设计视频生成或剪辑流程
+```
+
+## 一、结论
+
+在 GPT V2 改造后，文生图和图生图都属于 Codex / video-workbench 的阶段6以后执行层能力。
+
+GPT V2 只提供：
+
+- 视觉目标；
+- 画面功能；
+- 页面 / 镜头功能；
+- 人物、场景、风格、模板、图示系统；
+- 参考图需求；
+- 验收关注点。
+
+Codex 负责判断某个资产应走文生图还是图生图，并负责最终提示词、试跑、review、失败归因和版本落盘。
+
+核心规则：
+
+> 文生图用于从文字规格生成初始资产；图生图用于在已有参考资产上保持、变体、修正或扩展。
+
+## 二、文生图适用场景
+
+文生图适合没有强参考图依赖、可以通过文字规格清楚定义的资产。
+
+典型场景：
+
+- 科普视频的概念隐喻图；
+- 培训 PPT 的概念页背景；
+- 客户方案 PPT 的商务图示背景；
+- 信息图或流程图的初始视觉方向；
+- `visual-system/` 的第一轮风格试跑；
+- `slides/sNN/` 的第一版页面 / 镜头资产。
+
+输入来源：
+
+- GPT V2 阶段4通用表达与视觉系统；
+- GPT V2 阶段5页面 / 分镜 / 内容骨架；
+- `execution-plan.md` 中的当前批次策略；
+- `visual-system/visual-system.md` 中的当前接受版本；
+- 单元级 `sNN-unit-spec.md` 和 `sNN-visual-brief.md`。
+
+输出文件通常是：
+
+```text
+slides/sNN/sNN-v1-prompt.md
+slides/sNN/sNN-v1-image.png
+slides/sNN/sNN-v1-review.md
+```
+
+或在视觉系统试跑阶段：
+
+```text
+visual-system/iterations/<round-id>-prompt.md
+visual-system/iterations/<round-id>-image.png
+visual-system/iterations/<round-id>-review.md
+```
+
+## 三、图生图适用场景
+
+图生图适合已经有可用参考资产，并且新图需要保留某些视觉锚点时使用。
+
+典型场景：
+
+- 人物身份、服装、姿态、表情需要保持一致；
+- PPT 模板已经接受，需要生成同一体系下的其他页面背景；
+- 科普视频已经固定母题，需要生成后续镜头的同风格变体；
+- 某张图片构图基本正确，只需要修改局部问题；
+- 需要把外部平台生成的可用图转成项目统一风格；
+- 需要在不破坏核心构图的情况下修正文字、图标、密度、光线或主体位置。
+
+图生图的输入不应直接来自 GPT V2。它应来自项目内已经接受或待迭代的资产：
+
+```text
+visual-system/refs/
+visual-system/iterations/
+slides/sNN/sNN-vN-image.png
+```
+
+图生图提示词必须显式拆分：
+
+- must_preserve：必须保持什么；
+- must_change：必须修改什么；
+- can_change：可以自由变化什么；
+- must_avoid：必须避免什么；
+- reference_role：参考图在本轮中的作用；
+- acceptance_criteria：本轮通过标准。
+
+输出文件仍按单元版本命名：
+
+```text
+slides/sNN/sNN-v2-prompt.md
+slides/sNN/sNN-v2-image.png
+slides/sNN/sNN-v2-review.md
+```
+
+## 四、和 `visual-system/` 的关系
+
+`visual-system/` 是文生图和图生图的共同上游。
+
+推荐顺序：
+
+```text
+GPT V2 阶段4
+→ Codex 提炼 visual-system materialization targets
+→ 文生图生成第一轮公共风格 / 模板 / 角色 / 场景参考
+→ review
+→ 接受的资产进入 visual-system 当前接受版本或 refs
+→ 图生图用于保持一致性、扩展变体或修正问题
+```
+
+不要把图生图当成默认起点。没有稳定参考图时，图生图容易放大错误锚点。
+
+也不要把文生图当成全流程唯一手段。项目一旦有已接受视觉系统，后续需要一致性的资产应优先考虑图生图或参考图约束。
+
+## 五、和 `execution-plan.md` 的关系
+
+`execution-plan.md` 应记录当前批次使用文生图还是图生图。
+
+建议记录字段：
+
+```md
+## Current Round
+- round:
+- units_or_assets:
+- generation_mode: text-to-image | image-to-image | mixed
+- reference_assets:
+- prompts_to_generate:
+- images_to_generate:
+- review_focus:
+- next_action:
+```
+
+对于小批次试错，`execution-plan.md` 应说明本轮为什么选择这些单元：
+
+- 代表不同页面类型；
+- 覆盖高风险镜头；
+- 检查人物一致性；
+- 检查 PPT 模板复用；
+- 检查文字 / 图表承载策略；
+- 检查视觉母题是否能扩展。
+
+## 六、和 `slides/slides.md` 的关系
+
+`slides/slides.md` 只记录单元事实，不替代执行策略。
+
+建议在表格中记录生成模式：
+
+```md
+| unit | batch | mode | prompt | image | review | accepted | notes |
+| --- | --- | --- | --- | --- | --- | --- | --- |
+| s01 | batch-01 | text-to-image | v1 | v1 | pass | yes | establishes opening style |
+| s05 | batch-01 | image-to-image | v2 | v2 | revise | no | preserved style but composition too dense |
+```
+
+## 七、文字、公式、图表承载策略
+
+不再使用全局 `no readable text` 规则。
+
+Codex 在执行层判断信息由哪里承载：
+
+- PPT 层；
+- 页面 / 网页层；
+- 图片层；
+- 后期编辑层；
+- 图标、图示或抽象符号；
+- 其他工具。
+
+判断依据是媒介目标，而不是固定禁令：
+
+- 培训 PPT、科普 PPT、公式讲解、流程图、框架图可能需要文字、标签、公式或节点。
+- 科普视频的背景图、隐喻图、氛围图通常不应让图像模型生成正文。
+- MV / 剧情图通常更重视人物、场景、光线和情绪，不应塞入解释性文字。
+
+## 八、失败归因
+
+文生图常见失败：
+
+- 隐喻漂移；
+- 构图过散；
+- 页面留白不足；
+- 假文字或伪图表；
+- 风格和项目公共系统不一致；
+- 视觉资产像海报，不像可讲解页面组件。
+
+图生图常见失败：
+
+- 参考图锚点过强，改不动；
+- 参考图锚点过弱，一致性丢失；
+- 人物身份漂移；
+- 模板被破坏；
+- 局部修改造成整体构图退化；
+- 版本之间差异不可控。
+
+这些失败应记录到：
+
+```text
+visual-system/visual-system.md
+slides/sNN/sNN-vN-review.md
+```
+
+## 九、当前采用规则
+
+1. Codex 是最终图像提示词 owner。
+2. GPT V2 不输出最终文生图或图生图提示词。
+3. 文生图用于初始资产生成和低参考依赖资产。
+4. 图生图用于参考资产已经存在后的保持、变体、修正和扩展。
+5. `visual-system/` 先物化，单元资产再小批次推进。
+6. `execution-plan.md` 记录每轮采用文生图、图生图或 mixed。
+7. `slides/slides.md` 记录每个单元的实际生成模式和版本状态。
+8. 单元资产使用 `sNN-vN-type.ext` 命名。
+9. 是否让图片承载文字、公式、标签或图表，由执行层按媒介目标判断。
+10. 本调研不证明最终图片质量，只固化本地执行边界。
--- a/investigations/README.md
+++ b/investigations/README.md
@ -0,0 +1,20 @@
+# Investigations
+
+This directory holds research reports, experiments, comparisons, and draft workflow proposals for Video Workbench.
+
+Investigation files are not stable operating rules by themselves. Accepted conclusions must be promoted into `README.md`, `VIDEO_WORKBENCH.md`, `docs/`, or project templates before they become durable workflow rules.
+
+## Current Records
+
+```text
+2026-06-23-gpt-image-2-advisor-prompt-quality/
+2026-06-23-gpt-image-2-advisor-prompt-quality-v2-broader-use-cases/
+2026-06-23-text-to-image-image-to-image-execution-boundary/
+```
+
+Notes:
+
+- The first prompt-quality investigation over-weighted character and MV identity scenarios. It is retained as historical evidence.
+- The V2 prompt-quality investigation is the accepted basis for shifting final image prompt ownership to Codex.
+- The text-to-image / image-to-image execution-boundary investigation records local workflow policy only. It does not validate final image quality.
+- Any historical `makeup-still/` directory recommendation inside old investigation files is superseded by the current `visual-system/` rule.
--- a/tmp/.gitkeep
+++ b/tmp/.gitkeep
@ -1 +0,0 @@
-
Author	SHA1	Message	Date
wantsong	822b7456fe	docs: 完成 video workbench 重构案例库	2026-06-23 08:46:13 +08:00
wantsong	c7e208a99e	docs: archive video workflow investigations	2026-06-23 08:08:48 +08:00
wantsong	b12502bf8c	docs: align video workbench with GPT V2 workflow	2026-06-23 07:58:44 +08:00