learn

name	learn
description	/learn — extract patterns into skills from session history, and evaluate the session for memory-worthy content. NOT for writing skills from scratch (use wisdom). NOT auto-triggered by the word "learn" in a prompt — invoke explicitly or via the low-frequency memory_nudge hook (PreCompact/Stop).
when_to_use	learn from this session, save this pattern, extract reusable patterns, evaluate session for memory, session memory check
user-invocable	true

Launch the @learn agent (Task tool, subagent_type: learn) to analyze conversation history, evaluate it for memory-worthy content, and create or update skills.

Two capabilities, one skill

1. Memory evaluation (usually the low-effort pass, often nudge-triggered): review the session for corrections, confirmed decisions, project facts, or reference pointers and save them via the auto-memory mechanism (~/.claude/projects/<slug>/memory/, frontmatter name/description/metadata.type of user/feedback/project/ reference, indexed in MEMORY.md). This does not require user back-and-forth — save what clearly qualifies, skip what doesn't.
2. Pattern/skill extraction (heavier pass, always user-invoked): read conversation history, identify recurring themes, and propose new/updated skills for the user to approve (see agent process).

Run memory evaluation first when triggered by the nudge; run full extraction when the user explicitly asks to learn from a session.

Rules for extracted skills

• ALWAYS read the session transcript and identify the specific failure/decision being captured BEFORE drafting (path: see global skill startup protocol).
• NEVER promote a single-session story to a skill — need pattern in 2+ distinct sessions; otherwise record in .diary/.

Rules for memory evaluation

• ALWAYS distinguish the four memory types (user/feedback/project/reference) per the auto-memory format — don't dump everything into one bucket.
• NEVER save routine operations or one-off trivia — only durable, reusable facts or confirmed corrections.
• ALWAYS update MEMORY.md's one-line index when adding an entry.
• NEVER invoked by the bare word "learn" appearing in a user prompt — that keyword route was removed deliberately (see hooks/README.md). Only the slash command and the memory_nudge.py hook trigger it.