The Compound Engineering Flywheel

Compound engineering is the term Anthropic uses for the flywheel effect that emerges when skills, hooks, CLAUDE.md, and CI are all working together. Each PR that surfaces a new pattern adds a rule to CLAUDE.md. That rule improves every future session for every developer. Skills accumulate — each one is a permanently solved workflow. Hooks catch errors automatically — each one is a quality gate that never sleeps. CI reviews enforce standards consistently — each review comment is a potential new CLAUDE.md rule. The system gets smarter with every interaction.

The compounding is not linear — it is exponential. Week one, you have a basic CLAUDE.md and no skills. Claude makes the same mistakes everyone's Claude makes. Week four, you have twenty rules, five skills, and three hooks. Claude catches errors it used to miss, follows patterns it used to invent, and runs verification loops it used to skip. Week eight, your CLAUDE.md is a comprehensive operating manual, your skill library handles every daily workflow, and your hooks create an autonomous quality pipeline. New team members onboard in hours instead of days because the entire team's knowledge is encoded in files that Claude reads automatically.

The flywheel has a maintenance cost. CLAUDE.md needs pruning — contradictory rules and stale conventions accumulate. Skills need updating when workflows change. Hooks need testing when the build pipeline evolves. CI review criteria need refinement as the codebase grows. But the return on that maintenance investment is orders of magnitude larger than the cost. Every hour spent improving the system saves ten hours across the team over the following month. This is the endgame of Claude Code mastery: not writing better prompts, but building systems that make every prompt — from every developer — produce better results.

1. Layer 1: CLAUDE.md (Team Knowledge) Every mistake becomes a rule. Every convention becomes a standard. The file grows, gets pruned, and grows smarter. This is the institutional memory that every session inherits.
2. Layer 2: Skills (Automated Workflows) Every repeated workflow becomes a skill. Deploy, test, audit, scaffold — permanently solved. Each skill is a workflow that never needs to be re-explained.
3. Layer 3: Hooks (Autonomous Quality) Stop hooks run verification after every change. Post-tool hooks auto-format. Pre-execution hooks block dangerous commands. Quality enforcement that never takes a day off.
4. Layer 4: CI Reviews (Continuous Improvement) Claude reviews every PR against your CLAUDE.md standards. Issues become new rules. New rules improve future reviews. The feedback loop closes and the system self-improves.