my-skills — your skill collection, visualized

A meta-skill: one persistent registry of every skill you have, where each came from, who (which model, which chat) made it, and how they rank.

Engine: scripts/myskills.py (stdlib-only, fast). Persistent state: data/registry.json (atomic writes) + data/TABLE.md (auto-rendered markdown table, humans can open it directly). Alias: /msk.

Three surfaces, one engine

The registry has three doors — all views over the SAME engine and registry (one source of truth; every write goes through the same atomic save):

msk stats renders the aggregate dashboard (category/model/scope/origin bars, rating averages) straight in the terminal.

Standard flow (every invocation)

1. Always sync first — incremental and cheap after the first run:

   python3 ~/.claude/skills/my-skills/scripts/myskills.py sync

   It discovers new/changed/removed skills and attributes ONLY new ones (origin session, model). It prints what's new — surface that to the user.
2. Then render what was asked (default = full table, most recent first):

   python3 ~/.claude/skills/my-skills/scripts/myskills.py list

   ✨-marked green rows are NEW (added <48h ago or discovered this sync). Show the table output verbatim in a code block so alignment survives.

Command cookbook

Intelligent behaviors (you, the agent, do these)

• Natural-language filters/sorts — translate to flags above. If no flag fits, pull list --json and filter/sort/reason yourself, then present a numbered table matching the CLI table's columns.
• Ranking by arbitrary criteria ("rank by powerfulness / usefulness / impact on industry if published / creativity / …"): pull list --json, judge each skill from its description (Read its SKILL.md when the description is thin), score 1–10 with a one-line justification per skill, then PERSIST every score via rate <name> <criterion>=<score> --by <your exact model id> and render with --sort rating:<criterion>. Ratings are durable — future sessions build on them instead of re-judging.
• Questions about any skill ("what does X do", "is X safe", "which is better for Y") — info X, and Read its SKILL.md for depth. Never guess.
• "Which chats/sessions made my skills?" — origin data is already in the table; for one skill use sessions <name>. The origin transcript path in info lets you open the actual conversation for archaeology.
• Native/built-in skills — the engine can't see harness-native skills (code-review, loop, schedule, …). YOU can: they're listed in your system prompt. When the user wants them included, record what you actually see: add-native "code-review:Review the current diff…" "loop:Run a prompt on an interval" … then list --all. Ground truth from the harness, never a hardcoded list.
• Categorize better — sync auto-guesses categories by keyword; when the user asks for smarter grouping, re-bucket with set <name> category <cat> so it sticks.

Reading the table

• scope: global (~/.claude/skills) · proj:<dir> (project-local) · plugin:<name> · native (harness built-in, only with --all).
• origin: created@<session-title-or-id> — a Claude Code chat verifiably wrote it (earliest Write/Edit event in any transcript; evidence: shell means it was created via shell commands, slightly weaker). imported — it exists on disk but no transcript created it (copied in, git-cloned, or the origin chat was deleted; no session id to show, as expected).
• model / effort: the model that made the creating write, read from that very transcript event — never self-reported. Effort is best-effort from the session's cship snapshot (~/.cship/live/<sid>.json) and may be absent: Claude Code transcripts do not record effort.
• aliases are folded into their canonical skill (ep (+epp,epq) style), symlink-resolved — one source of truth, no duplicate rows.

Guarantees & safety

• Transcripts and skill dirs are read read-only; the engine writes ONLY under data/ (atomic temp+rename, no torn state).
• sync is incremental: attribution (the expensive grep over all transcripts) runs once per skill, single pass per candidate file, cached forever in the registry. Re-runs cost ~a second.
• Origins are evidence-backed: session id, timestamp, model, transcript path all come from the transcript itself and are kept for verification.