Teaching Claude Code to Learn from Mistakes

Boris Cherny created Claude Code. When he shared how his team uses it, one thing stood out:

"Anytime we see Claude do something incorrectly we add it to the CLAUDE.md, so Claude knows not to do it next time."

Simple idea. Every mistake becomes a rule. The project gets smarter over time.

But doing this manually is easy to skip. You fix the bug, move on, forget to document it. A week later, Claude makes the same mistake.

I built a plugin that makes this a single command.

The plugin

It's called learn. After Claude makes a mistake, you run:

/learn:from

Paste the error or describe what went wrong. The plugin investigates, proposes a rule, and adds it to your CLAUDE.md.

That's it.

Example

TypeScript throws an error after Claude edits a file. I run:

/learn:from

Claude asks what went wrong. I paste the error:

Type 'string' is not assignable to type 'number'.
  Property 'id' expected number but got string from API response.

Claude checks the recent changes, sees the pattern, and proposes:

Proposed rule:

> API responses are untyped at runtime. Always validate and transform 
> external data before assigning to typed variables.

Add to CLAUDE.md? (yes/no)

I type yes. The rule is added.

Next time Claude fetches from an API in this project, it knows to handle this.

Installation

/plugin marketplace add marcoshaber99/learn
/plugin install learn@marcoshaber99-learn

That's it. Works on any project with a CLAUDE.md file.

The idea

Every mistake becomes a rule. The longer you use Claude on a project, the fewer mistakes it makes.

Your CLAUDE.md becomes a knowledge base built from real failures, not hypothetical best practices.