Each unit of work should make the next unit easier - not harder. 80% plan & review, 20% execute & compound.
Plugin: EveryInc/compound-engineering-plugin · Version: v3.5.0 · Skills: 37 · Agents: 51
Run /ce-setup once per project, then cycle. Skip the last stage and you’ve built nothing reusable.
[1] IDEATE (optional) → /ce-ideate
│
▼
[2] BRAINSTORM → /ce-brainstorm
│
▼
[3] PLAN → /ce-plan
│
▼
[4] WORK → /ce-work
│
▼
[5] REVIEW → /ce-code-review
│
▼
[6] COMPOUND ★ → /ce-compound
│
└─ feeds back into [2] BRAINSTORM on next cycle
| Stage | What it produces |
|---|---|
| Ideate (optional) | requirements doc |
| Brainstorm | 3 research agents |
| Plan | worktrees + tasks |
| Work | parallel personas |
| Review | confidence-gated findings |
| Compound | docs/solutions/ |
v3.1: brainstorm now probes rigor gaps with prose before Phase 2 + product-tier ID traceability.
Each cycle starts smarter. Captured learnings auto-discovered next time.
Beyond the loop.
| Command | What it does |
|---|---|
/ce-strategy |
Anchor upstream of the loop. Captures problem, persona, metrics at STRATEGY.md; ideate / brainstorm / plan read it as grounding. v3.4. |
/ce-ideate |
Divergent + adversarial filter. Run when the idea is fuzzy. Ranks candidates and routes the strongest one into brainstorming. |
/ce-debug |
Root-cause hunter. Traces causal chains, forms hypotheses, test-first fixes. v3.2: now delegates commit/PR + branch check. |
/ce-optimize |
Iterative optimisation. Parallel experiments, measurement gates, LLM-as-judge scoring. Anything you can score numerically. |
/ce-doc-review |
Spec review via persona agents. Coherence, feasibility, scope-guardian, product/design/security lenses, adversarial. Run before /ce-plan. |
/ce-polish-beta |
HITL polish after review. Verifies CI, boots dev server from .claude/launch.json, dispatches polish sub-agents, emits stacked-PR seeds for oversized work. |
/ce-compound-refresh |
Stop learnings drifting. Audits docs/solutions/ - keep / update / replace / archive. Quarterly, or whenever a pattern contradicts. |
Stop hand-writing commits.
| Command | What it does |
|---|---|
/ce-commit (alias: /git-commit) |
Value-communicating commit message. Why, not what. |
/ce-commit-push-pr |
Commit → push → open or update PR. Delegates title + body to /ce-pr-description. Idempotent on re-runs. |
/ce-pr-description |
Regenerate value-first PR title and body from current branch. Reusable by other skills. |
/ce-resolve-pr-feedback |
Parallel feedback resolver. v3 tightens clustering to cross-round only, centralises test runs, drops bot noise. |
/ce-clean-gone-branches |
Prunes local branches whose remote tracking is gone. |
Beta · hands-off.
| Command | What it does |
|---|---|
/lfg |
Full autonomous chain. plan → deepen → work → review → resolve → browser → video → PR → compound. v3.1: ralph-loop dropped, ce-commit-push-pr added. Bounded scope only. |
/slfg |
/lfg with swarm mode. Parallel execution across work items. Trades coordination overhead for wall-clock speed. |
Utilities that earn their keep.
| Command | What it does |
|---|---|
/ce-riffrec-feedback |
Riffrec → structured feedback. 3-path routing: setup, quick bug report, or full analysis (hands off to /ce-brainstorm). v3.5. |
/ce-product-pulse |
Closes the loop. PM read on user outcomes - feeds back into strategy. v3.4. |
/ce-demo-reel (alias: /feature-video) |
Record demos for PRs. GIFs, terminal recordings, screenshots. Project-type-aware tiers. v3 adds local-save alt to catbox. |
/ce-test-browser |
Browser tests on PR-affected pages via agent-browser. v3.1: free-port scan + auto-server; pipeline mode skips headed prompt. |
/ce-sessions |
Cross-platform session history viewer. Resume across machines, mine prior runs for patterns. |
/ce-slack-research |
Slack archaeology via slack-researcher. Pulls decisions out of chat history. |
/ce-proof |
Create / edit / share docs via Proof editor. Reliable op shapes + retry/batch discipline. |
/onboarding |
Generate ONBOARDING.md from actual repo patterns - not boilerplate. |
/ce-update |
Plugin version check. Calls claude plugin update + diffs plugin.json on main, not tags. |
File-based backlog.
| Command | What it does |
|---|---|
/ce-todo-create |
Capture todo as a file in todos/. Survives session compaction - that’s the point versus in-chat TODOs. |
/todo-triage |
Prioritise pending todos across the folder. |
/todo-resolve |
Resolve todos in parallel - one work agent per item. |
Confidence-gated · parallel · headless.
kieran-rails · dhh-rails · kieran-python · kieran-typescript · swift-ios (new - SwiftUI state, retain cycles, Core Data threading, accessibility) · julik-frontend-races.
security-sentinel · performance-oracle · data-integrity-guardian · schema-drift-detector · architecture-strategist · api-contract · reliability · deployment-verification · adversarial · agent-native.
6-tier confidence rubric 0.00 - 1.00. The 0.60 threshold suppresses pre-existing issues, style nitpicks, and generic advice. Findings ranked P1 / P2 / P3.
- v3.2: tightened autofix (
safe_auto/gated_auto), LFG → best-judgment auto-resolve. - v3.3.1: reviewer queueing.
- v3.4: harness-native default, escalates on risk.
What the README won’t tell you.
- Run
/ce-setupfirst per project. v3.1:ast-grep; v3.4: codex global skills. - Reference
docs/solutions/fromCLAUDE.mdso agents auto-discover past learnings. - Scale-up order: build it → use it → trust it → then orchestrate.
/lfgis not step one. - Codex native install (v3.1+); Pi first-class via
pi-subagents. - Mix models per phase. Codex for implementation, Gemini for review - common Every setup.
- Compound while context is fresh - pattern over bug, before compaction eats specifics.
“Each unit of work should make the next unit easier. Remove yourself from as many places as you can - that’s where the compounding happens.”
— Every Inc · Klaassen, Shipper, Chow
Compound Engineering v3.5.0 reference card · field manual № 006 · prepared by Capita Technology.