itswenxin/motion-retrospective

Motion Retrospective

Every finished piece should leave the agent better than it found it. When a render is confirmed, review the whole build while the evidence is still in the runtime (processed commands are pruned after 7 days), and turn it into knowledge the next session can reuse.

Trigger

Start this skill in the same turn the user confirms a final render/export, for example: "渲染好了 / 导出了 / 就这版 / ship it / looks good, exporting". The bridge also adds bridgeHealth.retrospective to the result of any render command as a reminder. Do not wait to be asked again.

Steps

1. Collect. Note when the session started, then run:

   python3 tools/retrospective.py --slug <piece-name> --since <session start or e.g. 6h> [--match "<main comp name>"]
   

   This writes cases/_inbox/<date>-<slug>/ with TIMELINE.md, FAILURES.md, scripts/ (every successful mutating script, de-duplicated) and stats.json.

2. Capture evidence. Grab 4–6 frames from the final render (captureFrame or tools/run_capture.sh) and build a contact sheet with tools/sheet.py. Save into evidence/. Never commit the video, the .aep, or client footage and fonts.

3. Review the timeline. For each run of commands, write down:

   • what the user asked for (from the conversation), what changed, and which script did it;
   • the mechanism, not just the look (rig structure, expression, key values, layer order);
   • the numbers that made it work (durations, easing influence, offsets, sizes);
   • every failure in FAILURES.md: its cause and the fix.

4. Split into recipes. One piece usually holds several reusable techniques. Give each one its own folder cases/<category>/<technique>/ (categories: text-animation, shape-animation, ui-animation, transitions, effects) with RECIPE.md, a cleaned reusable .jsx, PROMPT_TEMPLATE.md and evidence. Keep a project index in cases/projects/<piece>/RECIPE.md that links them, like sync-dubbing-studio. Separate project-specific layer names from the transferable method.

5. Promote rules. Copy anything that applies beyond this piece into skill/SKILL.md: new pitfalls (坑位速查), new user-feedback translations (反馈翻译表), and the recipe index. Tick or add items in LEARNING_MAP.md. Update the case count and project table in README.md / README.zh-CN.md when they change.

6. Publish.

   tools/publish_case.sh cases/<category>/<technique> "what was learned" --push
   

   The script scrubs /Users/<name>/ paths, refuses video/audio/.aep files, runs tools/health_check.py, commits and pushes. Push automatically only if the repository owner opted in (MOTION_AGENT_AUTO_PUBLISH=1, or they asked for it); otherwise commit and ask.

7. Report to the user in a few lines: how many recipes were added, which rules were promoted, and the commit link.

Quality bar

• A recipe names the mechanism and gives numbers someone can reproduce.
• Every pitfall says what broke, why, and the fix.
• No client assets, private paths, or unreleased material in the public repository.
• If the piece taught nothing new, say so and add only the evidence to an existing case.