菜单
Mine past Claude Code session transcripts for recurring friction — problems the user had to find manually, corrections, repeated tool errors, hand-rolled loops — and turn the patterns into skill gotchas, memory entries, or CLAUDE.md rules. Use when asked to optimize the skills, run a retro, review past sessions, or find recurring problems.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Work off the Werkbank's Auftragskorb (the work_items basket) — read the open tasks, reproduce each complaint, restate it back into the row, triage it along the stage doctrine, fix the RULE rather than one spot, measure the effect and close the item with the diagnosed stage. Use when asked to work off the Korb, handle work items, process filed optimization tasks, or start an optimization round.
Take a finished change from diff to an open, green, review-clean PR — run the matching verify skills and the local CI gates first, then commit, push and open the PR, watch the pipeline AND the Copilot review, fix sensible findings and resolve the review threads. Never merges unless the user explicitly asks in this session. Use when asked to open or create a PR, ship a change, or finish up a change.
Run and exercise the FastAPI backend — start uvicorn, sweep the read endpoints with curl, check the diagnostic/fit/crop payloads against the live DB, and probe the admin write gate. Use when asked to run, start, test, or verify the API, a router, an endpoint, or a backend change.
Test and smoke-run the pure-Python compute layer in core/ — run pytest, run ruff lint/format the way CI does, and call the extraction pipeline directly on a synthetic chart without DB or HTTP. Use when asked to run, test, lint, or verify core/, the pipeline, extraction, template, or fit code.
Run and visually verify the kurrentschrift app after frontend changes — start the dev servers, click through ONLY the flows the change touches (chrome-devtools MCP locally, or playwright-core against the pre-installed Chromium in the cloud), watch the console and network, check style fidelity and legibility at a desktop AND a mobile viewport, and record a performance trace when the change can move performance. Use when asked to run, start, screenshot, verify, test in the browser, or visually check the app or a frontend change.
Verify Alembic schema migrations locally against a throwaway PostgreSQL — full upgrade chain incl. seeds, model↔migration drift via alembic check, and a downgrade/upgrade roundtrip — without ever touching the shared Cloud SQL DB. Use when asked to verify, test, or check a migration, a schema change, or alembic revisions before pushing.
| name | optimize-skills |
| description | Mine past Claude Code session transcripts for recurring friction — problems the user had to find manually, corrections, repeated tool errors, hand-rolled loops — and turn the patterns into skill gotchas, memory entries, or CLAUDE.md rules. Use when asked to optimize the skills, run a retro, review past sessions, or find recurring problems. |
Past sessions are data. This skill extracts friction signals from the
transcript JSONL files and converts recurring patterns into fixes in
the right place: a gotcha in an existing skill, a memory entry, or a
CLAUDE.md rule. Transcripts live under ~/.claude/projects/<slug>/,
where the slug is the absolute repo path with / replaced by - —
derive it from the working directory instead of hardcoding a
machine-specific path. The files are megabytes — never read them
raw, always extract with jq.
ls -S ~/.claude/projects/$(pwd | tr '/' '-')/*.jsonl
Four extractors, each verified against real transcripts. FILE is one
transcript path:
User texts — the user's own words: bug reports the assistant missed, corrections, re-instructions:
jq -r 'select(.type=="user") | .message.content | if type=="string" then . elif type=="array" then (map(select(.type=="text") | .text) | join(" ")) else empty end' FILE | grep -v '^$' | grep -v '^<' | head -80
Tool errors, deduplicated by class:
jq -r 'select(.type=="user") | .message.content[]? | select(type=="object" and .type=="tool_result" and .is_error==true) | .content | if type=="string" then . elif type=="array" then (map(.text? // empty) | join(" ")) else empty end' FILE | cut -c1-200 | sort | uniq -c | sort -rn | head -20
Interruptions (user broke in mid-action):
grep -c 'Request interrupted by user' FILE
Command repetition (hand-rolled loops a skill should own):
jq -r 'select(.type=="assistant") | .message.content[]? | select(type=="object" and .type=="tool_use" and .name=="Bash") | .input.command' FILE | awk '{print $1, $2}' | sort | uniq -c | sort -rn | head -10
Ignore: routine output, system reminders, skill texts pasted into user messages, one-off typos.
For 1–2 transcripts, run the extractors inline. For more, use a
Workflow: one mining agent per transcript (each runs the §1 extractors
and returns structured findings), then a single synthesis agent that
clusters across sessions, ranks by sessions × frequency, and —
after reading the current skill files — proposes only fixes that
aren't already documented. First run (2026-06-10): 13 sessions →
84 raw findings → 10 patterns.
Route each accepted pattern to where it prevents the recurrence:
verify-* / open-pr / write-docs skillApply small, low-risk fixes (gotchas, memory) directly and list them; anything touching settings, plugins, or new repo code is a recommendation for the user, not a unilateral change.
args can arrive undefined — inline the transcript
file list in the script body instead of passing it via args (a
run failed exactly this way on the first attempt)..message.content is string OR array depending on entry type —
both extractor branches are needed, otherwise jq silently drops
half the messages.grep -v '^<' filter drops system-reminder
blocks, the rest needs judgment in §2..type; inspect first:
jq -r '.type' FILE | sort | uniq -c.