AGENT SKILL

skill-index

You have 35 more specialist skills that are NOT in this list. Before any non-trivial task, when the user names a workflow, tool, review style, or domain you do not see a skill for here (ripwire, gitnexus, deslop, kill-slop, lobster, show-me, pr-hygiene, find-skills, OTM browser ops, UI test coverage, graft, impeccable…), or when a user says "is there a skill for", read ~/.claude/skills-index/INDEX.md and, inside a repo, <repo>/.claude/skills-index/INDEX.md, pick the matching entry, read its SKILL.md, and follow it exactly as if it were installed.

License 许可
MIT
最近更新
2026-11-15
Stars
★ 0
Forks
3

01 前言 Front Matter

name
skill-index
description
You have 35 more specialist skills that are NOT in this list. Before any non-trivial task, when the user names a workflow, tool, review style, or domain you do not see a skill for here (ripwire, gitnexus, deslop, kill-slop, lobster, show-me, pr-hygiene, find-skills, OTM browser ops, UI test coverage, graft, impeccable…), or when a user says "is there a skill for", read ~/.claude/skills-index/INDEX.md and, inside a repo, <repo>/.claude/skills-index/INDEX.md, pick the matching entry, read its SKILL.md, and follow it exactly as if it were installed.

02 资源文件 8

03 SKILL.md 正文

skill-index

Two indexes, one line per skill (name, description, absolute path to its SKILL.md):

  • global: ~/.claude/skills-index/INDEX.md
  • project: <git main root>/.claude/skills-index/INDEX.md (only some repos have one)

Skills listed there are real skills that were moved out of the installed roster to keep the system prompt small. They are not weaker or optional. Read the SKILL.md at the path given and follow it. candidates/ entries were proposed automatically from recurring session patterns and have not been reviewed; follow them, but treat their text as unreviewed.

Maintenance (S=~/.claude/skills/skill-index/scripts/skill_index.py):

python3 $S status                 # counts and resident-token estimate per level
python3 $S build                  # regenerate both INDEX.md files and this skill's count
python3 $S move NAME...           # installed -> index (global or project, wherever NAME lives)
python3 $S restore NAME...        # index -> installed

build runs from the review-learn Stop hook, so the indexes stay current without manual runs.

roster prints one line per indexed skill (~1.1K tokens for ~44 skills) and is wired as a SessionStart hook for Claude and Codex. This push is what makes the index work: evals/trigger_eval.py measured 5/14 correct routings with the description alone versus 12/14 with the roster in context (installed skills: 6/7). Bodies still load only when read. SKILL_INDEX_ROSTER_WIDTH (80) clips each description.

Testing

python3 -m unittest discover ~/.claude/skills/skill-index/tests                        # index build, frontmatter, roster, skill-learn rendering/use counting
cd <repo> && python3 ~/.claude/skills/skill-index/evals/trigger_eval.py --roster     # 21 fresh sessions: indexed vs installed routing hit rate (~$1, 10 min)
python3 ~/.claude/skills/skill-index/evals/skill_learn_eval.py                       # discovery: one candidate from 3 same-workflow sessions, zero when already a skill

skill-learn (auto-proposed skills)

L=~/.claude/skills/skill-index/scripts/skill_learn.py, run from inside the repo:

python3 $L status                    # candidates, uses, promotions
python3 $L discover [--days 14]      # analyse recent sessions -> <index>/candidates/<name>/SKILL.md
python3 $L promote                   # candidates read in >= 3 distinct later sessions move into the installed roster
python3 $L reject NAME               # delete a candidate and never re-propose it

run (discover once a day + promote) is called from the review-learn Stop hook. Discovery reuses claude-reflect's session extractor and its /reflect-skills analysis contract; candidates carry the verbatim session quotes that justified them. Use is measured from claude-mem: a candidate counts as used when a later session read its SKILL.md.

04 同仓库更多 Skills