# Style Anchor Trial Plan

## Status

- Document status: user accepted; superseded by style brief/spec/review artifacts.
- Execution status: planning completed; do not generate images from this file.
- Scope: common style anchor board for `在路上`.
- Next gated action after approval: completed through `style-anchor-brief-v1.md`, `style-anchor-board-spec-v1.md`, and `style-anchor-board-review-v1.md`.

## Purpose

The character anchor answers:

> Is this still the same traveler?

The style anchor answers:

> Do these images still belong to the same MV?

This trial should turn the four existing blueprint images into an execution-ready style system without mixing them with the character-anchor workflow.

## Source Images

Blueprint images:

| Anchor | Source file | Current role |
| --- | --- | --- |
| A | `intake/imgs/A.风景主导型蓝图.png` | Accepted base for landscape-led road composition. |
| B | `intake/imgs/B.高位观察型蓝图.png` | Accepted base for high-angle city-edge observation. |
| C | `intake/imgs/C.泥泞里的星空蓝图.png` | Problem reference; must be rebuilt darker with clearer broken starlight. |
| D | `intake/imgs/D.夜路篝火型蓝图.png` | Accepted base for restrained night-road fire warmth. |

Local contact sheet:

```text
visual-system/style/style-blueprints-contactsheet.png
```

## Source Files Actually Used

Primary project sources:

- `visual-system/visual-system.md`
- `intake/阶段四-通用表达与视觉系统设计.md`
- `intake/阶段四-风格锚点板页面结构建议稿.md`
- `project.md`
- `execution-plan.md`

## Initial Blueprint Assessment

### A: 风景主导型

Usable as a base anchor.

What works:

- long road;
- large world / small person;
- cold landscape-led road-film composition;
- low-saturation earth and gray palette.

Watch risk:

- light is somewhat sunrise-like. It is acceptable for A, but should not infect C.

### B: 高位观察型

Usable as a base anchor.

What works:

- city edge rather than city promotion;
- wet ground and rough terrain;
- person small enough;
- documentary observation feeling.

Watch risk:

- keep it away from clean skyline / tech-city / cyber UI drift in later prompts.

### C: 泥泞里的星空

Do not use as final C anchor.

What works:

- mud and puddle reality are present;
- low-angle puddle composition is directionally right.

Problem:

- too close to sunrise / sunset reflection;
- too warm and bright;
- broken star reflection is not clear enough;
- the spiritual function is not yet strong enough.

Required repair:

- dark night or pre-dawn;
- real muddy puddle as main subject;
- visible broken star reflection;
- low-place-sees-high-place feeling;
- no sunrise, no sunset, no fantasy magic, no shoe advertisement.

### D: 夜路篝火型

Usable as a base anchor.

What works:

- dark road;
- small fire;
- restrained warmth;
- distant faint lights;
- not a camping party or victory scene.

Watch risk:

- later D prompts must preserve solitude and distance, not become a gathering.

## Recommended Advancement

Do not directly make the final style board by simply placing the four current blueprints together.

Recommended sequence:

1. **Style brief**
   - Create `style-anchor-brief-v1.md`.
   - Record A/B/D as usable base references.
   - Record C as a repair target, not an accepted anchor.
   - Define the visual rules that the final board must enforce.

2. **C repair prompt/spec**
   - Create a Skill-based prompt document for C only.
   - Use `gpt-image-2` workflow and record mode, template/reference, source files, and input image roles.
   - Generate only after user approval.

3. **C repair generation/review**
   - Generate one C repair candidate.
   - Review it against the dark muddy starlight criteria.
   - If C fails, repair C before moving on.

4. **Style anchor board assembly**
   - Use accepted A, B, repaired C, and accepted D.
   - Prefer a deterministic board/layout artifact for readable labels and rules.
   - Do not rely on image generation to render exact Chinese labels.

5. **Five-shot test package**
   - Use the accepted character slices and style anchors in `S-01`, `S-10`, `S-46`, `S-47`, `S-51`.

## Board Construction Recommendation

For the final common style anchor board, prefer a layout/composition artifact rather than a single AI-generated board.

Reason:

- A/B/D already exist as useful references.
- C needs one targeted repair.
- The board needs exact Chinese labels and rules.
- Generated text inside an image is not reliable enough to carry execution rules.

Recommended board form:

- one 16:9 dark-gray board;
- four visual sample cells A/B/C/D;
- right-side compact rule column: color, light, composition, texture, drift avoidance;
- bottom review rule line;
- exact text handled by layout/document layer, not by image generation.

## Candidate Reusable Workflow Notes

Do not promote these as stable Video Workbench knowledge yet, but this trial can test:

1. whether style anchors should be split into accepted base anchors plus repair targets;
2. whether exact labels should be assembled outside image generation;
3. whether C-style soul images should always get a separate repair gate;
4. whether five-shot tests should only start after C is repaired.

## Next Gated Action

After this plan is approved, create:

```text
visual-system/style/style-anchor-brief-v1.md
```

That brief should be the project-facing source for:

- C repair prompt/spec;
- final style anchor board structure;
- later five-shot style references.