Second Brain

Compound Engineering: Only Document Mistakes

The Story

Boris maintains a file called CLAUDE.md that acts as living documentation for his AI workflow. But he doesn’t fill it with rules upfront. He only adds an entry when Claude actually fails repeatedly at the same task.

John Kim calls this “compound engineering: building institutional knowledge directly into the codebase so every session gets smarter.” Each failure teaches the system. Each entry prevents the same mistake from happening again.

The file stays small (roughly 2,500 tokens). It’s not a comprehensive manual. It’s a curated list of hard-won lessons.

Patricia Kimmerle points out that this is “fundamentally different from traditional documentation, which tries to be comprehensive upfront.” Boris’s approach is reactive, not preventive. Only real failures earn a spot.

Lesson for Creators

Build a system where every failure teaches the system, not just you. Keep a running document of things that went wrong and what you learned. Don’t try to anticipate every problem upfront. Let the failures come, capture the lesson, and your system gets smarter over time. This is how you compound knowledge instead of repeating mistakes.

Backlinks