Review Skills

Review discovered skills

The skill-extractor mines reusable procedures ("skills") from Claude Code conversation + coding traces, scores each by confidence (is this a real, generalizable skill?) and utility (how valuable if reused?), and queues the strong, novel ones for human review. This flow lets the user approve, edit, or reject them. Every candidate — approved or not — stays on record so mining improves over time.

The engine lives at ~/.claude/skill-extractor/. Use the review CLI for all actions (never hand-edit the state files). Resolve the paths first:

PY=$(cat ~/.claude/skill-extractor/extension-config.json | python3 -c "import json,sys;print(json.load(sys.stdin)['python'])")
ENGINE=$(cat ~/.claude/skill-extractor/extension-config.json | python3 -c "import json,sys;print(json.load(sys.stdin)['engineDir'])")

(Fallback: PY=python3, ENGINE=~/Nesh/skill-extractor/engine.)

Steps

1. List pending candidates: run "$PY" "$ENGINE/review.py" list. This returns

JSON with id, name, title, description, confidence, utility, composite, and trace_outcome for each. If empty, tell the user there's nothing to review and optionally offer to run the miner now ("$PY" "$ENGINE/extractor.py").

1. Present them to the user as a concise ranked list (highest composite

first): title, one-line description, and the scores like confidence 88% · utility 77% · trace: success. Briefly note what each score means the first time.

1. For the one(s) the user wants to look at, run

"$PY" "$ENGINE/review.py" show and show the full body (the procedure), trigger, and outcome_reason.

1. Take the user's decision for each candidate:

• Install (optionally with edits): if the user wants changes, write a JSON

file with only the changed fields (any of name, title, description, trigger, body, tags) to a temp path, then run "$PY" "$ENGINE/review.py" install --edits /tmp/edits.json --comment "". With no edits, drop --edits. This writes ~/.claude/skills//SKILL.md so it becomes a live skill.

• Reject: run

"$PY" "$ENGINE/review.py" reject --comment "". Always try to capture a short reason — it's the training signal for future mining. The candidate stays in the scratch store; it is not deleted.

1. Confirm what happened (installed path, or rejection recorded) and, if the

user installed a skill, remind them it's available immediately in new sessions.

Notes

• Never delete candidates. Reject keeps them; install promotes them. This is by design.
• If the user asks to see everything mined (not just pending), use list --all.
• The same actions are available via the VS Code popup ("Discovered Skills"

panel); this command is the terminal equivalent and stays in sync with it.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: surenode-ai
• Source: surenode-ai/skill-extractor
• License: Apache-2.0

Install and usage instructions live in the source repository linked above.