| name | cortex-learn |
| description | Detect routing patterns in the Cortex log and surface candidate Decision Shortcuts for cortex.md. Use when the user asks for routing patterns, wants to tune cortex defaults, or says "/cortex-learn". |
| trigger | /cortex-learn |
/cortex-learn
Run pattern detection on ~/.claude/cortex-log.jsonl. Surface groups of routing decisions that repeated ≥3 times as candidate Decision Shortcuts for cortex.md.
When to invoke
- User types
/cortex-learn explicitly
- User asks "what patterns is cortex seeing"
- User asks "are there routes I should codify"
- User says "tune cortex based on usage"
- Session-start check (via
cortex learn --check) surfaced a pending pattern
What to do
Step 1 — Run pattern detection
python3 ~/.claude/bin/cortex learn
Optional:
--threshold N — change minimum count (default 3)--check — terse one-liner for session-start use (no review flow)--no-proposals — skip writing markdown proposal files
Step 2 — Review the proposals
For each detected pattern, the tool writes a markdown proposal file with frontmatter status pending. The user should:
- Read each proposal
- Decide: approve, modify, or reject
- Edit the proposal frontmatter to reflect the decision
- If approved: add the candidate shortcut to
cortex.md → Decision Shortcuts table
- If rejected: write a one-line reason in the Notes section
Step 3 — Update cortex.md (only if approved)
Take the candidate shortcut row and merge it into the Decision Shortcuts table in cortex.md. Don't overwrite existing rows that already cover the same case.
Do NOT
- Do not auto-apply patterns to cortex.md — Amit must approve each one
- Do not delete proposal files even after approval; they're the audit trail
- Do not skip the
--no-proposals flag if the user explicitly asks for stats only
