Harvest

Harvest — turn repeated work into skill proposals

Scan the user's recent session transcripts, find work they keep re-doing or re-explaining, and draft each recurring pattern as a skill proposal. Proposals are drafts only — nothing becomes an active skill until the user approves it via /promote.

State directory

All Second Nature state lives in ~/.claude/second-nature/:

• proposals//SKILL.md — the drafted skill
• proposals//EVIDENCE.md — why it was proposed
• retired/ — skills retired by /curate
• LOG.md — append-only history of proposals, promotions, retirements

Create these directories on first run.

Steps

1. Collect recent history

Transcripts live in ~/.claude/projects//*.jsonl — one directory per project, one file per session, one JSON event per line.

• Default scope: sessions modified in the last 14 days, across all projects (or only the current project if the user says so). Find them with find ~/.claude/projects -name "*.jsonl" -mtime -14.
• These files can be very large. Do NOT read them whole. Extract only the signal:
• user prompt text (events whose type marks a user message)
• shell commands run (Bash tool inputs)
• skill/slash-command invocations

A grep/jq pipeline into a scratch file per session is the right shape. Sample long sessions rather than processing every line.

• If more than ~30 sessions match, prefer the most recent 30 and say so in the report — never silently truncate.

2. Detect candidate patterns

A pattern is skill-worthy when at least one of these holds:

• The same multi-step workflow appears in 2+ sessions (3+ is a strong signal).
• The user re-explains the same instructions or preferences to Claude more than once.
• A fiddly incantation recurs (long command lines, exact flag orders, API call shapes) that was clearly re-derived each time.
• A sequence ends in the user correcting Claude the same way repeatedly.

Ignore: one-offs, patterns already covered by an existing skill (check ~/.claude/skills/ and the project's .claude/skills/), and anything that is really a personal fact (that belongs in memory, not a skill — note it in the report instead).

3. Draft proposals

For each candidate, write ~/.claude/second-nature/proposals//:

SKILL.md — a complete, installable skill: frontmatter with name (short kebab-case) and description (what it does + concrete trigger phrases), then step-by-step instructions written from the observed pattern, generalized so it works next time — not a transcript replay.

EVIDENCE.md — the case for it: which sessions (session-id + date + project), how many occurrences, a one-line description of each occurrence, and an estimate of what the skill saves (time, tokens, or mistakes avoided).

Secrets rule (hard): transcripts may contain API keys, tokens, passwords, and private data. Never copy secret values into a proposal — reference them as placeholders ($API_KEY, ``). If a pattern is inseparable from private data, skip it and say why.

If a proposal for the same pattern already exists, update its EVIDENCE.md with the new occurrences instead of duplicating.

4. Report

Append one line per new/updated proposal to ~/.claude/second-nature/LOG.md (YYYY-MM-DD harvest: proposed (N occurrences)), then give the user a summary table: proposal name, what it automates, occurrences, sessions spanned. End by reminding them: review and install with /promote, nothing is active yet.

If nothing recurred enough to propose, say so plainly — an empty harvest is a valid result, not a failure.

Source & license

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

• Author: mmusa
• Source: mmusa/second-nature
• License: MIT

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