Compound Engineering Camp: Every Step, From Scratch — Katie Parrott

Parrott documents Cora general manager Kieran Klaassen’s live demonstration of compound engineering, his philosophy of software development where every bug, fix, and code review makes the AI system permanently smarter. The compound engineering plugin has 10,000+ GitHub stars and is used by engineers at Google and Amazon.

Key workflow steps: (1) Brainstorm — the plugin interviews you collaboratively, bridging vague idea to detailed spec. (2) Plan — researches codebase, checks existing patterns, surfaces past learnings, produces implementation plan with zero additional input. (3) Implement — executes the plan. (4) Compound — stores lessons as artifacts that future agents can discover. Klaassen emphasizes running this step while context is fresh, before conversation compaction loses specifics.

Model selection strategy: faster models (Claude Haiku 4.5, Gemini 2.5 Flash) for brainstorming, Opus for planning, Codex for implementation, sometimes Gemini for code review. Klaassen built a working app from a one-line prompt in under an hour, running 25 agents in parallel.

RDCO Mapping

Highly relevant to our own development workflow. The compound engineering pattern — persistent lessons across sessions, different models for different phases, review-before-commit — maps directly to how we should structure our build-project skill. The “compound when context is fresh” tip is especially actionable: we should add a post-build compounding step to capture what worked and what broke.