COMPOUND LEARNING

Most engineering accumulates debt. Every feature adds complexity. The codebase gets harder to work with over time. Compound engineering inverts this.

Every Inc's compound-engineering plugin for Claude Code ships with 24 specialized agents, 19 slash commands, and a four-phase workflow: Plan, Work, Review, Compound. The key insight is the ratio—80% planning and review, 20% execution.

The /compound command captures solutions as searchable documentation after each cycle. Plans inform future plans. Reviews catch more issues. Patterns get codified. The knowledge base from phase four feeds directly into phase one of the next cycle.

I've been running this workflow for three weeks. The difference shows up in velocity on the second and third iterations of similar tasks. Context that used to live in my head now lives in skills and agents that Claude can reference.

The plugin also converts to OpenCode and Codex formats via a Bun/TypeScript CLI. Cross-platform portability matters when you're building institutional knowledge—you don't want it locked to one tool.

Traditional development treats documentation as overhead. Compound engineering treats it as infrastructure. Every solved problem becomes a reusable pattern.

The teams that systematize learning loops will ship faster than teams that just ship. Raw output matters less than compounding output.