Methodology Pattern Extractor

Runs via context: fork — the fork has no memory of the invoking conversation, so it rescans meetings/*.md fresh each time.

When to invoke

• The user says "update methodology", "extract a pattern", "I feel this flow has happened several times now".
• ≥ 5 meeting logs have accumulated but methodology.md §Patterns is still empty.
• After hitting an important milestone (end of spec phase, v1 release, major refactor close-out).

Inputs

• All meetings/*.md
• Corresponding session JSONL (optional, for filling in details) — see

session-to-meeting-log's "Locating the session file" for finding the .jsonl by id (ls ~/.claude/projects/*/<sessionId>.jsonl) rather than computing the project directory; this skill has no [session-id] argument, so treat it as best-effort and skip it if the file isn't already known. Create custom subagents only documents this transcript location for subagent transcripts specifically — treat the general-conversation case as unconfirmed.

• The existing docs/methodology.md (to avoid duplicate entries)

Rule: ≥ 3 sightings or it's not a pattern

Don't turn a one-off behaviour into a pattern. Criteria:

• The same triggering situation appears in ≥ 3 different meeting logs
• The response action each time is consistent or convergent
• There is an observable outcome

Items with fewer than 3 sightings can go into methodology.md §Backlog as "candidate patterns", annotated with the sighting count.

Output format

One entry per pattern, five lines:

### <Pattern name>

- **Trigger**: <what context activates this pattern>
- **Action**: <the observed response action>
- **Outcome**: <the resulting outcome>
- **Next-time adjust**: <what you'd improve next time, may be empty>
- **Sightings**: <one date per meeting log, comma-separated — ≥ 3 different logs, not ≥ 3 mentions in one log>

Example:

### Section-by-section approval

- **Trigger**: User asks Leader to draft a document section (design.md / foundations.md / plan.md)
- **Action**: Leader advances one section at a time, waits for OK before moving on
- **Outcome**: Avoids the high cost of "whole-doc rejection and rewrite"
- **Next-time adjust**: Before dispatching a sub-agent, confirm all prerequisites for the section are Resolved
- **Sightings**: 2026-05-15, 2026-05-22, 2026-06-03

Procedure

1. Grep meeting logs for frequently appearing verbs / triggers ("dispatch", "review round", "prerequisite", "rejected", ...).
2. For each high-frequency trigger, revisit its context — does the same action recur?
3. Filter to candidates with ≥ 3 sightings.
4. Extract Trigger / Action / Outcome / Next-time / Sightings.
5. Append to methodology.md §Patterns; never overwrite existing entries.
6. If a candidate has only 1–2 sightings, add it to methodology.md §Backlog as a "candidate pattern".

Anti-patterns (also worth recording)

docs/methodology.md also carries a §Anti-patterns section (sibling of §Patterns) for practices tried but found unsuitable. The three sections have different entry bars:

Verification checklist

• Every pattern carries Sightings ≥ 3.
• No duplicates with existing entries (grep methodology.md first).
• Original meeting logs are never modified.
• Pattern names use kebab-case or short phrases, no more than 6 words.

Deviation considerations

• Solo project, single author: "recurrence" includes the author's own habits; the ≥ 3 rule still applies.
• Cross-project: this skill's default scope is one repo; if extracting across repos, explicitly cite each origin.

Related skills

• session-to-meeting-log: produces this skill's input.
• backlog-routing-by-topic: candidate patterns go into methodology.md §Backlog.
• subagent-review-cycles / spec-phase-orchestration / leader-developer-handoff-contract: codified forms of common patterns.