learn

name	learn
description	Use when you want to review the current session for skills that misfired (skipped, invoked-but-ignored, or triggered user corrections) and update them — scans the active transcript, classifies findings, and delegates fixes to /lets-make-a-skill with per-finding approval
user-invocable	true

Learn From This Session

Invoked at the end of (or partway through) a session. /learn reviews the active transcript, identifies skills that misfired, and delegates fixes to /lets-make-a-skill — with a per-finding approval gate between detection and edit.

First action: Invoke /listen with this argument:

/lets-make-a-skill update any skill that misfired in the current session, following the analysis playbook below.

/listen's enforcement layer guarantees /lets-make-a-skill is invoked for each approved finding. The playbook below is the work content /listen carries forward.

Analysis Playbook

1. Locate the transcript

Current session = most recently modified *.jsonl file in ~/.claude/projects/<slug>/, where <slug> is the current working directory with / replaced by -, then any leading - stripped.

If the directory is missing or the newest file is ambiguous (multiple files touched within the last second), ask the user to confirm or paste the session ID.

2. Scan for misfire signals

Review the transcript in two passes:

1. For each Skill tool-call in the transcript, inspect the invocation, its output, and the next ~3 user/assistant turns.
2. Also scan the surrounding user requests and Claude actions for places where a skill should have been invoked based on its description, but was never called at all — this is how Skipped findings surface, since by definition they leave no tool-call to iterate over.

Signal	What it looks like
Skipped	A skill's description clearly matched the task, but the skill was never invoked
Invoked-but-ignored	The skill was invoked, and Claude's next action directly contradicts its instructions
User correction	User writes "wait", "hold up", "hmm", "double check", "no", "don't", "stop", "that's wrong", "why did you…" within 3 turns of a skill
User invokes another skill to course-correct	User invokes another skill to double-check output, logic, consistency, security, actions, or tool use in this session
Discarded output	User reverts or overrides the artifact the skill produced
Hook-rejection loop	A blocked action is retried after rejection — a skill should have prevented the first try
Re-prompt	User restates or clarifies the same request — skill interpreted it wrongly

Also capture the rationalization verbatim — the sentence Claude used to justify the misfire. /lets-make-a-skill needs this to add a specific counter.

3. Classify each finding

• Misfire — the skill's text is wrong, incomplete, or its triggers don't match the scenario. Fixable via /lets-make-a-skill.
• Preference shift — the skill followed its text faithfully; the user simply wants different behavior going forward. Report separately; do not auto-update.

Decision rule: "would a careful reading of this skill have produced the user's preferred behavior?" Yes → preference shift. No → misfire.

4. Present findings

One row per finding:

skill · signal · evidence (line refs) · classification · proposed-fix

For each row ask: approve, skip, or reclassify.

5. Apply approved fixes

For each approved misfire, invoke /lets-make-a-skill with:

• the skill's file path,
• the observed rationalization/gap (verbatim),
• enough context for a pressure scenario.

/lets-make-a-skill carries its own TDD discipline — do not bypass it.

6. Update scope

Edits are allowed in:

• ~/.claude/skills/ (user skills)
• <cwd>/.claude/skills/ (project skills, if the directory exists)

Plugin-cache skills under ~/.claude/plugins/cache/** are read-only — plugin updates overwrite local edits. If a misfire lives there, surface the finding but do not edit.

What This Skill Does NOT Do

• Does not edit skills in the plugin cache.
• Does not apply fixes without per-finding approval.
• Does not analyze sessions other than the current one.
• Does not treat every user correction as a skill bug — preference shifts are filtered out and reported separately.
• Does not summarize or moralize about Claude's performance — output is a findings table, not a retrospective.

Edge Cases

• No skills used → report "no skill invocations found" and exit.
• All skills used correctly → report "no misfires detected".
• All findings are preference shifts → surface as a design-question list, do not call /lets-make-a-skill.
• Misfire lives in plugin cache → surface, note it cannot be edited, suggest filing upstream.
• Transcript unreadable / JSONL corrupt → report the error; do not proceed blindly.