brain_scan_transcripts

A read-only inference pass over your existing Claude Code session transcripts. Suggests candidate Standards (e.g. patterns where you corrected Claude with a non-obvious reason) but never auto-writes — you review and accept each one.

Parameters

Name	Type	Default	Description
sinceDays	number	30	Only consider transcripts modified in the last N days
maxCandidates	number	20	Cap on returned suggestions

Output

Found 3 candidate corrections from past transcripts:

  1. [home-matt-fleet] 2026-04-15T14:22:13Z
     "don't add Co-Authored-By to commits, Matt is the sole author"

  2. [home-matt-macpool] 2026-04-19T09:14:02Z
     "stop using sk_live_ keys for staging — use rk_test_ instead"

  3. [home-matt-fleet] 2026-04-22T16:33:51Z
     "never push directly to develop, always PR"

Review these and decide which to convert to Standards/. brain_remember writes the chosen ones.

What it looks for

Patterns in user message turns that suggest a correction or rule:

• “don’t”, “do not”, “stop”, “never”, “always”
• “I told you”, “you keep”, “you always”
• “that’s wrong”, “you got it wrong”
• “correct(ly)”

It returns the matching line + project + timestamp. It does not judge whether the pattern is a real Standard worth recording — that’s your call.

Why it’s opt-in

Transcripts are noisy. A regex pass would create dozens of low-quality notes that pollute brain_recall. Surfacing candidates for human review preserves signal-to-noise.

Recommended workflow

1. Run brain_scan_transcripts(sinceDays=90) after first install
2. Read the candidates; pick the 3-5 that are real cross-cutting standards
3. For each: brain_remember(category=standards, title=..., body=...)
4. Re-run quarterly to catch newly-emerged patterns

Privacy note

This tool reads ~/.claude/projects/* on the local filesystem. It does not send transcripts to any external service. The pattern matching is local regex.