improve

Improve

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.

This runs in a checkout of the repo, for the person who maintains it. A fix here reaches every user on their next plugin update. S="${CLAUDE_SKILL_DIR}/scripts". Run the scripts with python3 on Mac and Linux, py on Windows.

0. Who is running this

python3 "$S/improve.py" mode

maintainer only in a git checkout of the repo with AI_EDITOR_MAINTAINER=1: go on to step 1.

user: the plugin's own files are replaced on every update, so never edit them (no SKILL.md, reference, script or PRINCIPLES.md). Collect and pick as below, then save each accepted item in the user's own layer, which updates never touch (docs/customise.md in the repo): a setting with taste.py add ... --set; a word rule with python3 "${CLAUDE_PLUGIN_ROOT}/lib/ai_editor/custom.py" rule "<rule>"; a caption look as custom/captions/<name>.json plus taste.py set captions.preset <name>; template defaults as custom/templates/<type>.json. Then offer taste.py suggest so the maintainers can fix it for everyone. Stop there.

1. Collect

python3 "$S/improve.py" collect [--edits D] [--no-gh]

One line a rule, most repeated first: times it came in, from where (suggestion, taste regression, preview correction, review note, GitHub issue), the owner named, and up to three examples. A rule said by several users, or said again after it was applied, comes first.

2. Pick

Show the top of the queue and ask in the question box which to fix now (multi-select, the most repeated first, recommended). For each one, also ask when it is unclear: Every user (Recommended) / Only some creators' styles (it belongs in a style.json field, not a rule) / Not a fix (close it).

3. Fix each one, in this order

1. The rule. One line in ${CLAUDE_PLUGIN_ROOT}/PRINCIPLES.md under the right heading, naming

where it is enforced, and the same rule in the owning SKILL.md or reference. General, never about one video: "Zooms ease over at least 0.5 s", not "the zoom at 0:12 snapped".

1. The enforcement. The smallest code that makes the rule hold without anyone remembering it:

a FAIL in a check (check.py, quality.py, product.py check, ai_tells.py), a refusal in the script that would break it (gates.py), a default that cannot produce it, or a line in tests/check_plugins.py for a repo rule. Prefer tightening an existing check to adding one.

1. The test. An assert in that script's demo() (or tests/), built from a small synthetic

example of the failure. Run it before the fix and see it fail, then after and see it pass. A rule only an agent can follow (asking, explaining) gets an eval case in evals/ instead: prompt.md plus a tool_used or llm grader.

1. The record. docs/feedback-matrix.md: the rule, where it is written, what enforces it, what

tests it. Then the changelog line:

AI_EDITOR_MAINTAINER=1 python3 "$S/improve.py" log "quality.py: a zoom faster than 0.5 s is a FAIL (quality.py demo)"

4. Run everything

python3 tests/check_plugins.py
for p in . plugins/ai-editor plugins/creator-teardown; do claude plugin validate --strict "$p"; done

Then every demo in .github/workflows/check.yml, node tests/test_record.mjs, and when the renderer is set up tests/smoke.py and tests/golden.py. All green before handing over. Do not commit; the maintainer reviews the diff and the CHANGELOG.md entry and commits.

On a GitHub issue that is now fixed, offer to comment the changelog line and close it (gh issue close <n> --comment "..."), only after asking.

Never

• Never edit a user's taste.md, taste.json, rules.json or suggestions.jsonl: they are theirs.
• Never put personal data in the repo: no names, handles, emails, file names or footage from the

feedback. Write the general rule and a synthetic example. tests/check_plugins.py fails on known private names; read every example before it goes in.

• Never close the loop with a note alone. A rule with no enforcement and no test is not done.

Files

• scripts/improve.py: mode, collect, log (maintainer mode only), demo. Reads only, except CHANGELOG.md.
• ${CLAUDE_PLUGIN_ROOT}/lib/ai_editor/custom.py: the user's <library>/custom/ folder (user mode writes there).
• ${CLAUDE_PLUGIN_ROOT}/skills/taste/scripts/taste.py: where suggest and issue write the user side.