使用方法
- 复制安装命令，在自己的智能体环境安装前先检查源码包。
- 从原始文档核对所需工具、凭据和支持的宿主。
- 说明输入材料与期望交付物，并对照来源检查输出。
npx skills add https://github.com/nextwork-projects/ai-video-editor-cooked --skill skill-improver开始前需了解
- 安装不包含模型订阅、API 额度或访问文件的权限。
- 仓库星数和安装量反映活跃情况，不等于独立质量或安全测试。
按源仓库内容呈现，保留标题、案例、代码、表格、链接以及原文引用的演示图片。
Skill improver
Paths: ${CLAUDE_SKILL_DIR} means the folder containing this SKILL.md, and ${CLAUDE_PLUGIN_ROOT} the plugin folder two levels above it.
Every question to the user goes in the question box: call the AskUserQuestion tool (2-4 options, the recommended one first). Only in an agent without that tool, ask numbered questions in text.
taste saves one correction the moment the user says it. This skill works on a whole finished video: it reads every note, keeps the lasting ones and writes them into the style or skill that made the video.
K="${CLAUDE_PLUGIN_ROOT}/lib/ai_editor/skillrules.py", R="${CLAUDE_PLUGIN_ROOT}/lib/ai_editor/review.py", T="${CLAUDE_PLUGIN_ROOT}/skills/taste/scripts/taste.py". Run them with python3 on Mac and Linux, py on Windows.
Copy this checklist and tick it off:
- [ ] 1. Notes collected
- [ ] 2. One-offs left with the video
- [ ] 3. Target found
- [ ] 4. Each rule checked for a twin or a conflict
- [ ] 5. Lines shown, user said yes
- [ ] 6. Saved, log written1. Collect the notes
- Review-page notes and your replies:
edits/<name>/review.json(comments). After the final approve,wait's
LEARN: block already lists them.
- Preview fixes:
edits/<name>/corrections.jsonl. - Corrections the user typed in this chat about this video.
2. Keep only notes for every future video
A note about one moment stays with the video. A note about how edits should look becomes a rule.
| note | rule? |
|---|---|
| "cut the um at 0:12" | no: one moment |
| "the Notion card covers my chin" (said twice) | yes: "Cards sit above the chin, never over the face" |
| "captions are tiny again" | yes: "Captions one size bigger" (with set captions.size_pct when the settings table in taste maps) |
| "wrong logo for Jev" | no: one asset (the logo-pack skill fixes it) |
Write the behaviour, not the instance: "Cards fill the space they have", not "make the router card bigger".
3. Find the target
python3 "$K" target edits/<name>"kind": "skill"with apath: a skill style. Rules go into that SKILL.md under## Learned rules.
No path: the skill is not in the user's skill folders; say so and stop. A skill inside a plugin is never written (an update replaces it); save those rules with taste instead.
"kind": "settings": rules go to that style (stylepath) or, withstylenull, to My style, through
styleskill.learn (rules.json, style.overrides.json, and the style's own skill rebuilt).
4. Check for a twin or a conflict
Skill style: python3 "$K" similar <SKILL.md> "<rule>" lists the learned rules that look alike. Settings style: python3 "$T" report lists the rules saved so far.
- Same meaning, vaguer: sharpen the old line (
replace), never add a second one. - Opposite meaning ("logos bigger" vs a saved "logos small"): ask in the question box which wins, quoting
both: The new note (Recommended) / The saved rule. Keep only the winner (replace). Never keep both.
- Nothing alike: a new line.
5. Show the lines, then ask
Each rule is one line: the rule, Why: the reason, and the date. Example, note to line:
note: "the logo is on my face again" (rounds 1 and 2)
line: - Logos sit beside the head, never over the face. Why: two notes said a logo covered the chin. (2026-10-13)
file: ~/.claude/skills/my-edit/SKILL.md, ## Learned rulesShow every line exactly as it will be saved and the file it goes in. Then ask in the question box: Save these (Recommended) / Change one / Don't save. Save nothing before the yes.
6. Save
From the review page (the final approve's "Teach <style> from this video" panel): post the rules, and the boxes the user ticks are the yes:
python3 "$R" edits/<name> learn --propose '[{"rule": "Logos sit beside the head, never over the face", "section": "visuals", "why": "two notes said a logo covered the chin", "replace": "Logos beside the head"}]'
python3 "$R" edits/<name> learn --done # after `wait` prints the ticked ruleswhy and replace are optional; set (a taste setting) only for a settings style. --done writes a skill style's rules into its SKILL.md and a settings style's through styleskill.learn. A round's learn --scope this-style does the same for a skill style (the note is the reason).
Each saved rule shows on the review page as a receipt, "Saved to <style>: <rule>. Undo", and on the style's page under "Learned rules" with its day and the video it came from, where the user can Remove it. Undo and Remove take the line out again (taste.py forget, skillrules.py remove); never put it back by hand.
From the Skills page ("Improve with feedback", JOB skill-change): no video, only the user's words. Steps 3-6 the same; the target is the path= the job names. start/references/home.md has the status lines.
From the chat, after the yes:
python3 "$K" add <SKILL.md> "<rule>" --why "<reason>" [--replace "<old rule>"] --edit edits/<name> # skill style
python3 "$T" add <section> "<rule>" [--set key=value] --style <style.json> --edit edits/<name> # settings style
python3 "$T" add <section> "<rule>" [--set key=value] --edit edits/<name> # My styleadd prints duplicate when the same rule is already there (nothing written), sharpened when it replaced one, added otherwise. It also appends one line to ## Log at the end of the SKILL.md (the last 20 kept). A settings style's log is rules.json: taste.py report shows each rule's count and dates.
Tell the user in one line what was saved and where.
Files
${CLAUDE_PLUGIN_ROOT}/lib/ai_editor/skillrules.py:target,similar,add,remove,list,demo.${CLAUDE_PLUGIN_ROOT}/lib/ai_editor/review.py:learn --propose/--done(the Teach panel).${CLAUDE_PLUGIN_ROOT}/lib/ai_editor/styleskill.py:learn, a settings style's rules and skill.
