Learn

Distill knowledge into ONE reusable skill. No engine needed — use your existing tools (Read, WebFetch, Write) to gather, then author a SKILL.md.

<source>$ARGUMENTS</source>

Gather

Resolve <source> with existing tools:

• A directory/path → Read/Grep the relevant files.
• A URL → WebFetch it.
• "what we just did" / empty → distill the workflow from THIS conversation.
• Pasted text → use as-is.

Author (where to save)

Default: save to <project>/.claude/skills/<name>/SKILL.md (project-first, self-evolving). Add frontmatter metadata.origin: agent-learned. Use ~/.claude/skills/ only if the user explicitly says it is globally reusable. Put long scripts in scripts/, deep detail in references/.

Authoring Constitution (follow strictly)

1. Bias toward ACTIVE — a learn pass that produces nothing is a missed opportunity, not a neutral result. Find the durable lesson.
2. Preference order (pick the earliest that fits, to avoid skill-sprawl): (a) PATCH a currently-loaded skill → (b) UPDATE an existing umbrella skill → (c) ADD a references//templates//scripts/ file to one → (d) CREATE a new umbrella skill only if nothing fits.
3. Name at CLASS level — "how to do THIS CLASS of task". Forbidden: PR numbers, error strings, dates, "fix-X-today".
4. DO NOT capture: environment failures (missing binary, command-not-found), negative claims about tools ("X is broken" — these harden into months-long refusals), transient already-resolved errors, one-off task narratives. If a tool failed due to setup, capture the FIX, not "the tool is broken". DO capture systematic friction: recurring setup gaps that block multiple sessions (capture the fix + root cause), complexity bottlenecks where a task took far longer than expected (capture the pattern), undocumented toolchain steps. Threshold: if the same friction would hit a second person, capture it. Specificity required: "npm install fails on M1 due to missing Rosetta — fix: arch -x86_64 npm install" is capturable; "setup was painful" is not.
5. Description ≤ 60 chars — write it, then COUNT the characters; if over, cut it down. One sentence, ends with a period.
6. Strong signal = user correction ("stop doing X", "too verbose", "always do Y") → first-class skill material.

Output

Report: which skill was created/updated, its path, its trigger description, and why. Concise.