Learn

Learn — Capture Knowledge into CLAUDE.md

You extract lessons learned from the current conversation and persist them into the project's CLAUDE.md file so that future sessions benefit.

How It Works

The user invokes /learn [optional topic] after discovering something important — a gotcha, a pattern, a convention, a debugging insight. You analyze the conversation, extract the key learnings, and append them to CLAUDE.md in the right section.

If a topic is provided, focus on that area. If not, scan the entire conversation for learnable moments.

Process

Step 1 — Identify Learnings

Review the current conversation for:

• Corrections the user made — "no, don't do it that way" moments
• Surprising behaviors — things that weren't obvious from reading the code
• Discovered conventions — patterns that should be followed going forward
• Gotchas and pitfalls — mistakes that future sessions should avoid
• Architecture decisions — why something was done a certain way
• Tool/library quirks — non-obvious behavior of dependencies

Step 2 — Read Current CLAUDE.md

Read the project's CLAUDE.md to:

• Understand the existing structure and sections
• Check if the learning is already documented (avoid duplicates)
• Find the right section to add the new knowledge

Step 3 — Draft the Addition

Write the learning in the style and format of the existing CLAUDE.md:

• Match the tone (terse rules vs detailed explanations)
• Match the formatting (bullet points, tables, code blocks)
• Keep it concise — future Claude sessions will read this every time
• Focus on actionable rules, not narrative

Step 4 — Present to User

Show the user what you want to add and where:

I'd add this to the [Section Name] section of CLAUDE.md:

> - **Rule/Convention:** Explanation of what to do or avoid

Wait for the user's approval before editing.

Step 5 — Apply

After approval, edit CLAUDE.md with the addition.

What Makes a Good CLAUDE.md Entry

Good entries are:

• Actionable — tells future Claude what to DO or NOT DO
• Concise — one rule, one line if possible
• Non-obvious — wouldn't be inferred from reading the code
• Durable — will be true for a long time, not ephemeral

Bad entries (skip these):

• Things already in the code (function names, file paths that can be found by reading)
• Temporary state ("currently working on X")
• Things that are obvious from the framework/language

Rules

• Never edit CLAUDE.md without showing the user first — always present the draft addition
• Don't duplicate — if the learning is already documented, say so
• Keep it minimal — add the smallest useful addition, not a paragraph
• Match existing style — read CLAUDE.md first and follow its conventions
• All text in English — CLAUDE.md content must be in English