video-workbench/projects/2026-06-23-在路上/slides/broad-generation-plan-v1.md

75 lines
3.0 KiB
Markdown

# Broad Generation Plan V1
## Status
- Document status: paused pending repair.
- Scope: lightweight broad production for all missing shot images after the accepted five-shot visual trial.
- User decision: skip per-shot review gates and generate missing shots in sequence.
- Existing accepted shots must be skipped: `s01`, `s10`, `s46`, `s47`, `s51`.
- Pause reason: self-audit found that the broad-run process removed too much traceability and did not actually use character reference images for visible character shots.
See:
```text
slides/broad-generation-self-audit-v1.md
```
## Goal
Generate the remaining `not started` shot images from `slides/slides.md` while preserving the accepted character anchor, common style anchor board, and five accepted core-shot results.
## Non-Negotiables
1. All project-bound image prompts still use the `gpt-image-2` Skill workflow.
2. The current mode is Skill Mode B: render prompts through the Skill workflow, then use the host image tool.
3. Existing accepted images are not regenerated unless the user explicitly asks for repair.
4. Each generated image is saved under its own `slides/sNN/` directory.
5. No per-shot human review gate is required during this broad run.
## Lightweight Artifact Pattern
Use one batch-level run log instead of one full review package per shot:
- `slides/broad-generation-run-v1.md`: mode check, shared source basis, queue, prompt/result trace.
- `slides/sNN/sNN-v1-image.png`: generated image.
- `slides/sNN/sNN-v1-prompt.md`: only when the shot needs a dedicated prompt record beyond the batch log.
No `sNN-v1-review.md` is required for broad-run shots unless a repair is requested.
## Status Rules
| State | Meaning |
| --- | --- |
| `not started` | No image generated yet. |
| `generated` | Image generated in the broad run, awaiting batch-level user review. |
| `accepted` | User accepted the image or the existing core-shot image. |
| `repair-needed` | User flagged the shot for local repair. |
The broad run should mark new images as `generated`, not `accepted`. User acceptance happens after batch-level review.
## Execution Loop
1. Read `slides/slides.md`.
2. Build the queue from rows whose status is `not started` and whose `sNN-v1-image.png` does not exist.
3. For each queued shot, render a compact prompt from:
- `visual-system/visual-system.md`;
- the shot row in `slides/slides.md`;
- accepted character/style anchors;
- nearby accepted core shots when relevant.
4. Generate one image with the host image tool.
5. Copy the selected output to `slides/sNN/sNN-v1-image.png`.
6. Append prompt/result details to `slides/broad-generation-run-v1.md`.
7. Update the row status in `slides/slides.md` to `generated`.
8. Continue to the next queued shot.
## Final Batch Review
After the queue finishes, create a compact batch summary:
- generated count;
- skipped accepted shots;
- missing or failed shots;
- optional contact sheet if useful.
The user reviews the batch result once and requests targeted repairs only for failed or drifting shots.