Compound Engineering Guide: Claude Code's 7-Step Loop

Key takeaways

1. Canonical definition. Compound engineering is the approach where every unit of engineering work makes subsequent units easier, not harder. Reversal of the complexity trajectory: bug fixes eliminate entire categories of future bugs, patterns become tools, the codebase becomes easier over time.
2. Main 7-step loop. Ideate → Brainstorm → Plan → Work → Review → Polish → Compound → (Repeat). Three phases: human decision (what to build?), agent execution (plan/code/test/PR), human judgment (is the output sufficient? did the system learn?). 1. Ideate (/ce-ideate): ambiguity → prioritized product options (confidence/impact/complexity scoring). 2. Brainstorm (/ce-brainstorm): vague idea → concrete requirements, saved in docs/brainstorms/. 3. Plan (/ce-plan): 3 parallel research agents (repo, framework-docs, best-practices) + spec-flow-analyzer → structured plan in docs/plans/. 4. Work (/ce-work): isolation (git worktree), step-by-step execution, validations, PR. 5. Review (/ce-code-review): specialized reviewers in parallel, findings prioritized P1/P2/P3. 6. Polish (/ce-polish-beta): actually clicking through the app, what "feels wrong" (speed, animation, copy, empty states). 7. Compound (/ce-compound): "the critical step" — capture the solution in docs/solutions/ with YAML frontmatter, update CLAUDE.md, create agents. "Would the system catch this automatically next time?"
3. The plugin. (EveryInc/compound-engineering-plugin): 40+ specialized agents, 30+ slash commands, 35+ skills. Review = ~23 reviewers (always-on: correctness, testing, maintainability, project-standards, agent-native, learnings; conditional: security, performance, API contract, migrations, reliability, adversarial; stack-specific: Rails, Python, TypeScript, Swift/iOS, frontend-races). Installable on Claude Code, Cursor, Codex (+ Copilot, Factory Droid, Gemini, etc.).
4. Project tree. CLAUDE.md/AGENTS.md (most important file, read every session) + .compound-engineering/config.local.yaml + docs/brainstorms/, docs/plans/, docs/solutions/ (categorized: performance-issues, security-issues, architecture-patterns, conventions…).
5. /lfg. autonomous development from idea to merged PR (Plan → Work → review+autofix → browser tests → commit/push/PR).
6. 50/50 rule (revision of the 80/20). 50% of time building features, 50% improving the system (review agents, pattern docs, test generators). ROI: 1 hour spent creating a review agent saves 10 hours over the year — system improvements compound, features don't.
7. 8 beliefs to unlearn. code must be hand-written / every line manually reviewed / the solution comes from the engineer / code is the primary artifact / first drafts must be good ("first attempts have a 95% garbage rate") / code is self-expression ("the code is not really yours") / the more you type the more you learn ("understanding matters more than muscle memory") / engineering thinking ≠ product thinking.
8. 8 beliefs to adopt. extracting one's taste into the system; the 50/50 rule; trusting the process + safety nets; the agent-native environment ("if a developer can see or do something, the agent should too"); parallelization (the new bottleneck is compute, not human attention); "Plans are the new code"; "engineers are product people now"; letting agents run while away.
9. 5-stage adoption scale. 0 manual → 1 chat-assistance → 2 agentic with line-by-line review (most plateau here) → 3 plan-first, PR-level review (key transition: compound engineering starts here) → 4 idea→PR (1 machine) → 5 parallel cloud execution (multi-feature, multi-device).
10. Agent-native — 4 levels. (1) basic dev (files, tests, commits); (2) full local (browser, logs, PR); (3) prod visibility (read-only logs, error tracking); (4) full integration (tickets, deployment) → unlocks stage 5.
11. Skip permissions. alias cc='claude --dangerously-skip-permissions'; the name is deliberately scary; safety nets = git reset, tests, PR review, worktrees. 5-10× gain in flow. Never in production or while learning.
12. 3 questions without tooling. "What was the hardest decision?" / "What alternatives did you reject and why?" / "What are you least confident about?"
13. Detailed best practices. disposable baby app to iterate on design; figma-design-sync + design-iterator; codifying design/copy taste into skills; structured user research (personas, frontmatter) readable by AI; usage pattern extraction (heavy/struggle/workaround/abandonment); copy = UX (codifying voice); automated product marketing (release notes + Playwright screenshots from the plan).
14. Strong link to prior coverage. consolidation text of the philosophy disseminated by [[shipper-klaassen-compound-engineering-every-agents-2025-12-11]], [[klaassen-stop-coding-start-planning-every-2025-11-06]], [[klaassen-teach-ai-think-senior-engineer-every-2025-11-07]]; v260 review pipeline [[chow-compound-engineering-v260-review-pipeline-2026-03-31]]; level 5 "Workflows / harness" from [[taylor-entis-every-eight-levels-ai-adoption-2026-06-02]].