video-workbench/projects/2026-06-23-在路上/slides/s46/s46-v1-prompt.md

10 KiB
Raw Blame History

S-46 Shot Prompt V1

Status

  • Document status: created for direct generation.
  • Execution status: used once for Mode B host image generation; candidate image accepted by user.
  • Project: 在路上.
  • Unit: s46.
  • Source shot: S-46.
  • Segment / time: Final four lines / 3:58-4:16.
  • Style anchor: C泥泞里的星空.
  • Purpose: generate the third trial image for the muddy starlight soul shot.
  • Generated candidate: s46-v1-image.png.

User Gate

User accepted S-10 and requested the next shot.

The current trial workflow skips intermediate prompt review. The user reviews final generated images.

Skill Mode Check

Command:

node C:\Users\wangq\.agents\skills\gpt-image-2\scripts\check-mode.js --json

Result:

{
  "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)",
  "summary": "MODE B / C · 未启用 Garden。若宿主 Agent 自带图像工具image_generation / dalle / mcp__*image* 等)→ MODE B把 prompt 交给宿主出图。若宿主无图像工具 → MODE C仅产出高质量 prompt 给用户。"
}

Interpretation:

  • Use gpt-image-2 as the prompt-engineering workflow.
  • Garden local generation is not enabled.
  • This Codex environment has a host image-generation path, so execution is Mode B.
  • Do not call Garden scripts for this step.

Skill Files Actually Read

  • C:\Users\wangq\.agents\skills\gpt-image-2\SKILL.md
  • C:\Users\wangq\.agents\skills\gpt-image-2\references\prompt-writing.md
  • C:\Users\wangq\.agents\skills\gpt-image-2\references\scenes-and-illustrations\concept-scene.md
  • C:\Users\wangq\.agents\skills\gpt-image-2\references\scenes-and-illustrations\minimalist-mood-scene.md
  • C:\Users\wangq\.agents\skills\gpt-image-2\references\product-visuals\premium-studio-product.md

Template adaptation:

  • concept-scene.md is used only for foreground / subject / background layering.
  • minimalist-mood-scene.md is used for restraint, quiet mood, and avoiding overfilled spectacle.
  • premium-studio-product.md is used as a negative reference: the boot must not become a product or advertisement subject.
  • Final output should be a realistic cinematic documentary still, not concept art, not product photography, and not fantasy illustration.

Source Files Actually Used

Primary shot source:

  • slides/s46/s46-v1-brief.md

Execution control:

  • project.md
  • execution-plan.md
  • slides/slides.md
  • slides/five-shot-trial-plan-v1.md

Visual system:

  • visual-system/visual-system.md
  • visual-system/characters/character-identity-lock.md
  • visual-system/style/style-anchor-board-review-v1.md
  • visual-system/style/style-anchor-c-repair-review-v1.md

Intake:

  • intake/阶段五-镜头清单骨架.md
  • intake/阶段五-核心单镜.md

Reference Assets

Primary C reference:

visual-system/style/style-anchor-c-repair-v1.png

Role:

  • primary C repair reference;
  • mood / composition / texture / acceptance reference;
  • not an edit target.

Style board reference:

visual-system/style/style-anchor-board-v1.png

Role:

  • common style anchor reference;
  • use for palette, road-film grammar, and drift controls.

Character / boot references:

visual-system/characters/character-anchor-v2-codex-posture-test.png
visual-system/characters/character-slice-walk-side-v2-from-anchor.png

Role:

  • peripheral boot / coat / lower-body identity reference;
  • not portrait target;
  • not edit target.

S-46 Target

This image must answer:

泥泞里,有星空。

The viewer should first see:

  • real mud;
  • dark water;
  • grit, stones, shoe marks, uneven puddle edge;
  • then discover broken star reflection inside the puddle.

The image should not show the character fully. The traveler is present only as an edge trace.

Rendered Prompt

Use case: photorealistic-natural.

Asset type: 16:9 MV trial shot still for 在路上, S-46, final-four-lines muddy starlight soul shot.

Primary request:

Generate one 16:9 horizontal realistic cinematic documentary still for the S-46 muddy starlight shot of the 在路上 MV. This is the soul image: a real muddy puddle at night or deep pre-dawn, with broken star points reflected in dark water. The image must say that the mud remains, but meaning appears inside it.

Scene:

A close-to-ground view of an irregular muddy puddle on a rough wet road or muddy ground at night or deep pre-dawn. The scene is dark, cold, and physically real. The puddle is surrounded by heavy wet mud, grit, stones, shoe marks, broken puddle edges, dirty water, small ripples, and uneven muddy texture. The water surface reflects a broken starry sky: small cold-white star points and faint scattered clusters interrupted by mud, ripples, stones, and footmarks.

Main subject:

The main subject is the muddy puddle and the broken star reflection in the dark water. The viewer should first read wet mud and real ground, then discover the stars inside the puddle. The star reflection must be visible enough to read, but it should be broken, partial, and physically interrupted by mud and water texture.

Peripheral traveler trace:

Only a small edge trace of the traveler is allowed: an old worn muddy boot at the far edge, a partial lower-leg shadow, or a small coat-hem edge. This trace should occupy under about 15% of the frame. It must not show the full person, face, portrait, or hero pose. The boot is old, road-worn, dirty, and practical, not new, not polished, not a fashion boot, and not a product closeup.

Composition:

Low ground-level composition. The camera is close to the mud. The puddle occupies the center or lower-center of the frame, with rough mud surrounding it. The boot or lower-body trace stays on the extreme edge only, cropped and secondary. Do not make the real sky the main image. The stars should be read as a reflection in puddle water, not as a large clean sky scene.

Style:

Real documentary texture + restrained poetic realism + low-place-sees-high-place composition. Low saturation. The image should belong with the accepted C repair anchor: dark blue-gray, wet black-gray, mud brown, coal black, and small cold-white star points. It should feel physically muddy and quiet, not decorative and not magical.

Lighting:

Night or deep pre-dawn darkness. Weak cold reflected light on water. Small cold-white star glints in the puddle. Dark and midtone areas should retain mud and water texture. No visible sun. No sunrise. No sunset. No warm horizon. No broad orange or gold glow. No firelight. No theatrical fantasy light.

Mood:

Quiet, low, difficult, and meaningful. The emotion is not victory and not escape. The mud is still there. The star reflection is discovered inside the mud.

Strict constraints:

  • 16:9 horizontal frame;
  • close-to-ground composition;
  • muddy puddle as main subject;
  • broken starry sky reflected in dark water;
  • realistic wet mud texture;
  • real grit, stones, shoe marks, water ripples, and dirty puddle edges;
  • old worn boot or lower-body trace only at the frame edge;
  • boot / traveler trace under about 15% of the frame;
  • night or deep pre-dawn darkness;
  • poetic documentary realism;
  • no portrait emphasis;
  • no full person;
  • no text, no labels, no logo, no UI, no watermark.

Strictly avoid:

  • sunrise;
  • sunset;
  • broad warm horizon;
  • beautiful postcard reflection;
  • clean mirror lake;
  • fantasy portal;
  • magic glow;
  • cosmic spectacle;
  • Milky Way as a real sky above the scene;
  • shoe advertisement;
  • product photography;
  • clean new boot;
  • fashion outdoor gear;
  • full-body traveler;
  • face or portrait;
  • heroic pose;
  • only mud with no stars;
  • only stars with no mud;
  • overly saturated colors;
  • text, symbols, watermarks, UI.

Final image behavior:

The viewer should first see real wet mud and dirty water. Then the broken star reflection should become clear inside the puddle. The boot or traveler trace should only prove this belongs to the traveler on the road, while the true subject remains mud, water, and broken starlight.

Reference Handling At Generation Time

Current execution:

  • Use the rendered prompt as a standalone prompt.
  • The local reference assets were read / viewed for prompt construction.
  • The host image tool path does not provide explicit file-reference arguments in this run.

If a later host path supports image references:

  • attach visual-system/style/style-anchor-c-repair-v1.png as primary C reference;
  • attach visual-system/style/style-anchor-board-v1.png as global style reference;
  • attach visual-system/characters/character-slice-walk-side-v2-from-anchor.png only as boot / lower-body reference;
  • explicitly instruct the model not to copy a full character and not to make the boot the subject.

Generation Result

Generated after user authorization to continue directly:

slides/s46/s46-v1-image.png

Review file:

slides/s46/s46-v1-review.md

Generation mode:

gpt-image-2 Skill Mode B; host image-generation tool; no Garden script.

Acceptance Criteria

S-46 passes only if:

  1. The viewer first sees real mud and water.
  2. Broken star reflection is clear enough to read.
  3. The scene is night or deep pre-dawn, not sunrise or sunset.
  4. The boot / traveler trace stays peripheral and under about 15%.
  5. The boot does not become a product shot.
  6. The image is restrained and documentary-real, not fantasy.
  7. It truly conveys "泥泞里,有星空".

Failure Conditions

Reject the generation if any of these happen:

  • no clear star reflection appears;
  • mud looks fake, clean, or decorative;
  • the image becomes sunrise or sunset;
  • broad warm light appears;
  • the reflection becomes a clean pretty landscape;
  • the image becomes fantasy, portal, or magic;
  • boot becomes the main subject or shoe advertisement;
  • a full person, face, or portrait appears;
  • the shot no longer feels connected to the accepted C repair.

Review File Created After Generation

After generation, created:

slides/s46/s46-v1-review.md

That review compares the generated image against:

  • slides/s46/s46-v1-brief.md;
  • this prompt/spec;
  • accepted C repair reference;
  • accepted common style anchor board;
  • accepted v2 character identity lock for boot/lower-body cues;
  • the S-46 acceptance criteria above.