weeder

!"${CLAUDE_SKILL_DIR}/bin/weeder" mode --banner

Agents that do not expand the line above (e.g. Codex): run <this-skill-dir>/bin/weeder mode --banner first and follow it.

The description loads every session; the SKILL.md body loads on every trigger. Never optimize on token count alone: a shorter SKILL.md that routes worse or drops an instruction is a regression. Every step below produces before/after evidence. Run all commands as python3 <this-skill-dir>/scripts/weeder.py <cmd>.

1. Audit: audit <skill_dir>. Token counts per section (description vs. body, chars/4 estimates unless calibrated), duplicate rules, filler, movable background/examples. Read all of it; optimize acts on only part.
2. Mechanical pass: optimize <skill_dir> --out <skill_dir>-optimized. Moves background/examples sections into references/ with a pointer. Lossless and reversible, so it's safe to always apply; it does not touch the description or duplicates.
3. Judgment pass, in the -optimized copy: shorten the description (keep concrete trigger nouns/verbs), state each duplicated rule once where the workflow needs it, cut only instructions that are genuinely obvious to a capable agent; don't cut something just because it's short. Optional structured help: suggest <skill_dir> --assist slight --json, then apply-suggestion <skill_dir> --answer <answer.json> --out <skill_dir>-optimized (see references/llm-assist.md; an assisted answer gets no free pass).
4. Routing: test-routing <skill_dir> --examples <examples.json> [--competing <other_skill_dir> ...] on both original and copy. examples.json = {"positive": [...], "negative": [...], "ambiguous": [{"prompt":, "expected": "trigger"|"no_trigger"|"either"}]}; if none exists, write 5-10 real triggers, a few unrelated prompts and borderline cases. Always pass sibling skills as --competing. Limits of this lexical proxy: references/routing-heuristic.md.
5. Function: test-function <skill_dir> <skill_dir>-optimized checks every never/must/do-not sentence survives (in SKILL.md or a reference). A MISSING constraint is a hard blocker, not a warning.
6. Report: diff <skill_dir> <skill_dir>-optimized --examples <examples.json> --competing ....
7. Accept only if routing accuracy after >= before (minus a small tolerance: a couple of points of a lexical proxy is noise) AND constraint preservation is 100%. Otherwise do another pass in step 3; don't lower the bar to make a bad result pass.

With the plugin, prefer delegating the whole workflow to the weeder agent (it preloads this skill and returns only the report) so the audit output stays out of your context.

Classification rules and why step 3 isn't automated: references/classification.md.