Compound Engineering

Workflow — the primary skill per SDLC stage this framework runs, top to bottom (folded and off-stage steps omitted). Validate and Review are sibling gates that both run after Implement.

flowchart TB
    A["ce-brainstorm<br/><i>Align</i>"] --> P["ce-plan<br/><i>Plan</i>"] --> I["ce-work<br/><i>Implement</i>"]
    I --> V["ce-dogfood<br/><i>Validate</i>"]
    I --> R["ce-code-review<br/><i>Review</i>"]
    V --> Rel["ce-commit-push-pr<br/><i>Release</i>"]
    R --> Rel
    Rel --> L["ce-compound<br/><i>Learn</i>"]

Compound Engineering is Every’s methodology (and open-source Claude Code plugin) for coding when AI agents write ~100% of the code. Its one-line thesis: “each unit of engineering work should make subsequent units easier — not harder.” Where traditional development accumulates technical debt, Compound Engineering accumulates knowledge — every bug, failed test, and problem-solving insight is captured and fed back so future agents avoid the mistake. The distinctive move is a learning loop: complexity grows alongside accumulated AI knowledge, so velocity accelerates over time rather than decaying. Authors Dan Shipper & Kieran Klaassen report single developers at Every now run five products — “the work of five developers a few years ago.” MIT-licensed; portable across Claude Code, Cursor, Codex, Copilot, Factory Droid, Kimi, Qwen, OpenCode, Pi, and Antigravity.

Install (Claude Code): /plugin marketplace add EveryInc/compound-engineering-plugin then /plugin install compound-engineering.
Effort allocation: ~80% planning and review, ~20% work and compounding — front-load thinking, and the execution + capture are cheap. (ce-plan alone is ~40% of a developer’s time.)

Distinctive contribution: the seventh stage

Compound Engineering is the framework that promotes a new canonical SDLC stage in this wiki: stage-learn. Its /ce-compound step — “the money step… the whole point” — captures reusable learnings into a docs/solutions/ corpus that future ce-brainstorm/ce-plan/ ce-code-review runs read as grounding. That gave learning-capture a second framework (alongside BMAD’s bmad-retrospective), clearing the bar to split it out of stage-release. See pattern-knowledge-compounding and stage-learn. Arguably the first genuinely new stage the agent era adds — the loop-closing arrow, not a rename of a classic phase.

Where gsd is an end-to-end workflow engine and bmad is a persona-oriented pipeline, Compound Engineering is a loop that feeds itself: it is the only framework here whose lifecycle is explicitly circular, with a dedicated step whose product is consumed by the front of the next iteration.

The six-step loop (+ autonomous mode)

Step	Command	What it does	Canonical stage
Brainstorm	`/ce-brainstorm` (ce-brainstorm)	Define what something should become through dialogue → requirements-only plan (`docs/brainstorms/`)	stage-align
Plan	`/ce-plan` (ce-plan)	Bound execution with guardrails; U-IDs + test scenarios; WHAT not HOW (`docs/plans/`)	stage-plan
Work	`/ce-work` (ce-work)	Execute against guardrails; figure out HOW with code, in an isolated worktree	stage-implement
Simplify	`/ce-simplify-code` (ce-simplify-code)	Refine fresh code for reuse/quality/efficiency, preserving behavior	stage-review
Review	`/ce-code-review` (ce-code-review)	Multi-agent persona review, confidence-gated, four modes	stage-review
Compound	`/ce-compound` (ce-compound)	Capture learnings → `docs/solutions/`; next iteration starts smarter	stage-learn
(all, hands-off)	`/lfg` (lfg)	Autonomous full pipeline: plan → work → simplify → review+fix → test → commit → push → PR → CI-green	stage-plan…stage-release

Distinctive mechanisms

Knowledge compounding as a first-class step — ce-compound + ce-compound-refresh maintain docs/solutions/ (artifact-solution-doc); a research council of sub-agents mines completed work, and learnings-researcher re-injects it downstream. This is pattern-knowledge-compounding, the framework’s signature.
80/20 front-loading — planning and review dominate; /ce-work is the cheap middle. Guardrails written in ce-plan bound what /ce-work may do (WHAT-not-HOW split).
Autonomous full loop — lfg chains every step hands-off and self-corrects until CI is green (pattern-autonomous-loop).
Worktree isolation — ce-worktree / ce-work run each unit in a dedicated git worktree so parallel and experimental work never corrupts the mainline (pattern-worktree-isolation).
Persona-based, confidence-gated review — ce-code-review and ce-doc-review fan out skill-local reviewer personas (pattern-parallel-persona-review) with adversarial and cross-model passes (pattern-adversarial-review), reporting only findings above a confidence bar.
Everything is grounded — ideation, briefs, and verdicts cite dual evidence (repo + web) and read the compound corpus first (pattern-source-grounding, pattern-context-engineering).

Capabilities

Skills — Align / discovery

ce-strategy — create/maintain STRATEGY.md, the upstream anchor read by ideate/brainstorm/plan.
ce-ideate — discover strong, qualified directions with conceptual frames + adversarial filtering.
ce-pov — decisive Adopt/Trial/Hold/Reject verdict on external inputs, dual-grounded.
ce-brainstorm — define what to build through dialogue → requirements-only unified plan.
ce-sweep — ingest Slack/GitHub items since cursors, acknowledge, analyze, reconcile plans (feedback intake).

Skills — Plan

ce-plan — enrich requirements into implementation-ready guardrails (U-IDs, test scenarios).
ce-doc-review — review requirements/plans with persona lenses before any code.

Skills — Implement

ce-work — execute the plan; figure out HOW with code; ship through quality gates.
ce-worktree — ensure work happens in an isolated git worktree.
ce-debug — systematic root-cause via causal chains, predictions, post-fix polish.
ce-polish — conversational UX polish with a live dev server + browser.

Skills — Review (quality gate; CE loop steps 4–5)

ce-simplify-code — reduce complexity in fresh code while preserving behavior.
ce-code-review — multi-persona, confidence-gated code review, four modes.
ce-optimize — metric-driven iterative optimization with parallel experiments.

Skills — Validate (functional testing)

ce-test-browser — end-to-end browser tests on PR/branch-affected pages.
ce-test-xcode — build/test iOS apps on simulator via XcodeBuildMCP.
ce-dogfood — hands-off diff-scoped browser QA with autonomous small-breakage fixes.

Skills — Release

ce-commit — one well-crafted atomic commit with convention awareness.
ce-commit-push-pr — working changes → open PR, three workflow modes.
ce-resolve-pr-feedback — evaluate/fix/reply to PR review feedback in parallel.
ce-promote — draft user-facing launch announcements across channels.
ce-product-pulse — time-windowed post-release usage/perf/error report.

Skills — Learn (the compounding step)

ce-compound — capture reusable learnings into docs/solutions/ (the money step).
ce-compound-refresh — maintain the solution corpus over time.
ce-explain — dense visual explainers of concepts/diffs/recent work (personal learning).

Orchestrator

lfg — the full hands-off pipeline from planning through green PR.

Sub-agents (26, each paged)

Compound Engineering skills delegate to a large shared cast of sub-agents (each skill’s references/agents/; many, e.g. ce-repo-profiler and ce-learnings-researcher, are dispatched by several skills). Each is paged as a subtype: sub-agent capability, filed under the stage of its intrinsic activity (grouped below by function).

Research / grounding:

ce-repo-profiler — question-agnostic cached project profile (stage-align).
ce-learnings-researcher — re-injects the docs/solutions/ corpus into new work (stage-align).
ce-slack-researcher — org-knowledge digest from Slack (stage-align).
ce-web-researcher — external grounding digest for ideation (stage-align).
ce-issue-intelligence-analyst — theme-level issue-tracker intelligence (stage-align).
ce-external-evidence-researcher — verified external evidence for verdicts (stage-align).
ce-precedent-activity-scout — prior-decision precedent + incumbent pain (stage-align).
ce-project-grounding-scout — this-codebase floor for verdicts (stage-align).
ce-media-analyzer — feedback media → bug-shaped finding (stage-align).
ce-git-history-analyzer — repo archaeology for planning (stage-plan).
ce-session-historian — mine past agent sessions for lessons (stage-learn).
ce-best-practices-researcher — authoritative best-practice synthesis (stage-learn).
ce-framework-docs-researcher — framework/version docs for lessons (stage-learn).
ce-work-recap-scout — gather recent-work evidence for explainers (stage-learn).
ce-repo-research-analyst — optimization-input repo research (stage-review).

Review / quality specialists:

ce-security-sentinel — attacker-minded security audit (stage-review; counterpart to addy-security-auditor).
ce-performance-oracle — bottleneck analysis (stage-review; counterpart to addy-web-performance-auditor).
ce-pattern-recognition-specialist — design/anti-pattern + recurring-problem analysis (stage-review).
ce-data-integrity-guardian — data-safety / migration-integrity review (stage-review).
ce-data-migration-reviewer — migration correctness + rollback, plan-time (stage-plan).
ce-deployment-verification-agent — launch-readiness checklists, plan-time (stage-plan).
ce-pr-comment-resolver — implements one validated PR-review fix (stage-release).

Design / planning:

ce-architecture-strategist — architecture-alignment review (stage-plan).
ce-agent-native-planning-strategist — decide agents-as-users, feed plan inputs (stage-plan).
ce-spec-flow-analyzer — surface missing flows/edge-cases pre-implementation (stage-plan).
ce-figma-design-sync — pixel-perfect Figma↔code sync (stage-implement).

The ce-session-historian + ce-pattern-recognition-specialist + ce-security-sentinel + ce-performance-oracle + ce-best-practices-researcher + ce-framework-docs-researcher + ce-data-integrity-guardian council is what powers ce-compound’s mining of completed work into solution docs.

Review persona lenses (bundled within review skills)

ce-code-review carries ~16 persona lenses (correctness, security, performance, maintainability, reliability, testing, api-contract, data-migration, adversarial, agent-native, deployment-verification, project-standards, previous-comments, learnings-researcher, + Swift/iOS and frontend-races stack reviewers).
ce-doc-review carries 7 (coherence, feasibility, security-lens, product-lens, design-lens, scope-guardian, adversarial-document).

Infrastructure / product-specific (catalogued, not paged)

/ce-setup — diagnose optional tool capabilities and bootstrap project-local config (infra bootstrap, like GSD/Addy setup skills — not a lifecycle stage).
/ce-proof — publish/view/comment/edit markdown via Every’s Proof collaborative editor (product integration).
/ce-riffrec-feedback-analysis — structure raw Riffrec recordings into feedback (Every-product integration; feeds ce-sweep/stage-align).

Artifacts produced

artifact-brainstorm-md — docs/brainstorms/ requirements-only unified plan (ce-brainstorm).
artifact-strategy-md — STRATEGY.md, the upstream strategic anchor (ce-strategy).
artifact-plan-md — docs/plans/ implementation-ready guardrail plan (ce-plan).
artifact-review-report — persona review findings (ce-code-review, ce-doc-review).
artifact-atomic-commit — one atomic commit per unit (ce-commit).
artifact-pull-request — the opened PR (ce-commit-push-pr, lfg).
artifact-solution-doc — signature — docs/solutions/ reusable learning (ce-compound, ce-compound-refresh).
artifact-explainer — dense visual explainer of a concept/diff (ce-explain).

Patterns applied

pattern-knowledge-compounding — signature — each unit of work leaves reusable, machine-consumable lessons future agents auto-consume (ce-compound, ce-compound-refresh).
pattern-autonomous-loop — chain the full pipeline hands-off with self-correction to a terminal success gate (lfg).
pattern-worktree-isolation — isolate each unit of work in a dedicated git worktree (ce-worktree, ce-work).
pattern-parallel-persona-review — concurrent reviewer-persona fan-out (ce-code-review, ce-doc-review).
pattern-adversarial-review — adversarial + cross-model review passes (ce-code-review).
pattern-plan-verification-loop — gate the plan/requirements before execution (ce-doc-review).
pattern-fresh-context-subagents — stateless persona/scout sub-agents per task.
pattern-context-engineering — read the compound corpus + repo profile as grounding.
pattern-source-grounding — dual-grounded ideas/verdicts (ce-ideate, ce-pov).
pattern-measure-first — metric-driven optimization (ce-optimize).
pattern-systematic-debugging — causal-chain root-cause (ce-debug).
pattern-trunk-based-development — atomic commits, PR flow (ce-commit, ce-commit-push-pr).

SDLC Agent-Skill Wiki

Explorer

compound-engineering