Learn

How it works

The /learn skill turns a source into a clean, reusable NanoClaw skill. Point it at a directory, a URL, a block of pasted notes, or the work you just did together, and it produces a new .claude/skills/<name>/SKILL.md — plus optional scripts/, references/, or templates/ folders — authored to the project’s skill guidelines.

It is instruction-only. There is no separate distillation engine and no reach-ins into core code; the skill works entirely through tools Claude already has: Read, Grep, Glob, WebFetch, and Write. That makes the output easy to inspect and easy to maintain.

Use it when you want to capture a workflow, not import one. /learn creates or refines skills from a source — it does not install existing community skills from a registry.

What it does

The skill runs through a fixed procedure:

1. Identify the source — a path is read as code, a URL is fetched, “what we just did” uses the current conversation, and pasted text is used directly. It also checks .claude/skills/ for an existing skill on the same topic. If one exists, this becomes a refine rather than a fresh create.
2. Gather the material — Glob the structure and Read the key files for a path, WebFetch a URL, or re-read what was actually done in the conversation. The goal is to understand the repeatable procedure, not every line.
3. Distill the procedure — strip the one-off specifics and keep the repeatable shape: the trigger, the step-by-step commands and file paths, and the gotchas that were hit along the way.
4. Author the SKILL.md — write the required name and description frontmatter, then an opening paragraph, a ## When to use section, and a ## Workflow of numbered steps. When refining, it reads the current file first and weaves new learnings in without duplicating steps or overwriting what is still correct.
5. Place and verify — write into .claude/skills/<name>/, confirm the folder name matches the name frontmatter and the YAML parses, dry-run the procedure where feasible, and tell you how to invoke it (/<name>).

Tips

• The most valuable content is the gotchas, not the happy path — capture the non-obvious things that went wrong and how you got past them.
• Keep each skill focused on one capability, mirroring the project’s “one change per PR” rule.
• Prefer instruction-only skills. If Claude can do the job by following prose plus existing tools, ship no code — those are the easiest to maintain and merge.
• When a source proves an existing skill stale — a changed path, command, or flag — re-run /learn on it to update the skill in place.
• description is what the agent reads to decide relevance, so make it concrete and include the phrases a user would actually say.