Skip to content
/

Compound Engineering

compound-engineering · EveryInc/compound-engineering-plugin · ★ 17k · last commit 2026-05-25

Make each unit of engineering work compound into easier future work via brainstorm→plan→execute→review→learn cycles.

Best when80% of agent effort should be in planning and review; knowledge accumulation (compounding) is a first-class workflow phase, not an afterthought.
Skip ifImplementing before brainstorming, Editing plan body during execution
vs seeds
bmad-methodin its persona-md subagent architecture and plan-first philosophy, but diverges in four ways: (1) adds a compounding mem…
Primitive shape 81 total
Commands 1 Skills 38 Subagents 42
00

Summary

Compound Engineering — Summary

Compound Engineering is a cross-runtime AI coding plugin (17,199 stars) created by Every Inc. that operationalizes the philosophy that "80% of effort belongs in planning and review, 20% in execution." The plugin ships 38 skills (user-invocable slash commands) and 50+ specialized subagent personas, covering the complete cycle from strategy to idea brainstorming to plan generation, implementation with worktrees, multi-persona code review, and knowledge compounding. Its most distinctive architectural trait is the "compounding" concept: a dedicated /ce-compound skill persists team learnings as durable COMPOUND notes in the repo so that future agents access accumulated institutional knowledge rather than rediscovering it. Multi-agent code review (/ce-code-review) spawns 10–20 tiered reviewer personas in parallel, each producing structured JSON findings that are merged and deduplicated by the orchestrating skill. Cross-runtime portability is first-class: the same plugin installs on Claude Code, OpenAI Codex, Cursor, GitHub Copilot, Factory Droid, and Qwen Code via a shared plugin manifest with per-runtime adapter scripts. Compared to seeds, it is closest to BMAD-METHOD (personas-as-agents, plan-then-execute cycle, skill-based entry points) but diverges by targeting multi-model cross-review (using RepoPrompt/Codex/Copilot as secondary reviewers), by introducing the compounding/knowledge-accumulation loop as a first-class workflow phase, and by supporting the widest cross-runtime installation surface of any framework in this corpus.

01

Overview

Compound Engineering — Overview

Origin

Created by Kieran Klaassen at Every Inc. (every.to), a technology media company. The plugin emerged from Every's internal AI-assisted development workflow and was open-sourced as a reference implementation of "compound engineering" — a term coined in the accompanying essay "Compound Engineering: How Every Codes With Agents" (every.to/chain-of-thought/...).

Philosophy

"Each unit of engineering work should make subsequent units easier — not harder."

Traditional development accumulates technical debt. Compound Engineering inverts this by structuring each cycle to leave behind reusable artifacts: brainstorm docs, plan docs, review findings, and COMPOUND notes that future agents read as grounding. The ratio is explicit: 80% planning and review, 20% execution.

"A good brainstorm makes the plan sharper. A good plan makes execution smaller. A good review catches the pattern, not just the bug. A good compound note means the next agent does not have to learn the same lesson from scratch."

The Core Loop

ce-brainstorm (WHAT) → ce-plan (HOW) → ce-work (DO) → ce-code-review (VERIFY) → ce-compound (LEARN)

ce-strategy is upstream of all cycles as the durable product anchor. ce-product-pulse closes the feedback loop with real user signal.

Target Audience

Engineering teams at scale who need structured handoffs between AI sessions. The plugin is opinionated about plan-first development: every feature should have a brainstorm doc before a plan, and a plan before execution.

Version

v3.8.4 (at analysis date 2026-05-25)

02

Architecture

Compound Engineering — Architecture

Distribution

  • Type: Claude Code plugin (also Codex, Cursor, Copilot, Factory Droid, Qwen Code)
  • npm package: @every-env/compound-plugin
  • Plugin manifest: plugins/compound-engineering/.claude-plugin/plugin.json
  • Version: 3.8.4

Installation Methods

Runtime Method
Claude Code /plugin marketplace add EveryInc/compound-engineering-plugin then /plugin install compound-engineering
Codex codex plugin marketplace add EveryInc/compound-engineering-plugin + bun script for agents + TUI install
Cursor /add-plugin compound-engineering in Agent chat
GitHub Copilot VS Code command palette Chat: Install Plugin from Source
Factory Droid droid plugin marketplace add ... + droid plugin install ...
Qwen Code qwen extensions install EveryInc/compound-engineering-plugin:compound-engineering

Directory Tree

EveryInc/compound-engineering-plugin/
├── .claude-plugin/marketplace.json         # Marketplace registration
├── .agents/plugins/                        # Global agent mounts
├── .claude/commands/triage-prs.md         # Single global command
├── .cursor-plugin/                         # Cursor-specific plugin dir
├── plugins/
│   └── compound-engineering/
│       ├── .claude-plugin/plugin.json     # Plugin manifest (v3.8.4)
│       ├── .codex-plugin/                 # Codex-specific
│       ├── .cursor-plugin/                # Cursor-specific
│       ├── agents/                        # 50+ agent persona files (.md)
│       ├── skills/                        # 38 skill directories (SKILL.md)
│       ├── AGENTS.md                      # Codex/Copilot agent manifest
│       └── CLAUDE.md                      # Claude Code instructions
├── docs/skills/                           # Extended skill documentation
├── scripts/                               # Install/cleanup scripts
├── src/index.ts                           # CLI entry point (bun)
└── package.json                           # @every-env/compound-plugin

Required Runtime

  • Claude Code, Codex CLI, Cursor, or Copilot (any one)
  • Node.js / Bun for the CLI installer (Codex-specific path only)
  • No external services required

Target AI Tools

Claude Code (primary), OpenAI Codex, Cursor, GitHub Copilot CLI, Factory Droid, Qwen Code

Config Files

  • plugins/compound-engineering/.claude-plugin/plugin.json — plugin identity
  • CLAUDE.md — project instructions
  • AGENTS.md — Codex/Copilot agent declarations
  • .flow/launch.json (optional, dev-server launch config)
03

Components

Compound Engineering — Components

Skills (38 total, user-invocable as slash commands)

Core Workflow

Skill Purpose
ce-strategy Create/maintain STRATEGY.md — product anchor read by all planning skills
ce-ideate Big-picture ideation: generate + critically evaluate grounded ideas
ce-brainstorm Interactive Q&A → right-sized requirements doc (WHAT to build)
ce-plan Turn feature ideas into detailed implementation plans (HOW to build)
ce-work Execute plans with worktrees, task tracking, test-first support
ce-debug Systematic failure reproduction, causal chain tracing, test-first fixes
ce-compound Document solved problems to compound team knowledge
ce-compound-refresh Refresh stale learnings (keep/update/replace/archive)
ce-optimize Iterative optimization loops with parallel experiments + LLM-as-judge
ce-product-pulse Time-windowed usage/error/performance report saved to docs/pulse-reports/

Code Review

Skill Purpose
ce-code-review Multi-agent code review: tiered personas, confidence gating, dedup pipeline
ce-doc-review Document review with parallel persona agents
ce-simplify-code Parallel reviewers find issues → fixes applied → behavior verified

Git Workflow

Skill Purpose
ce-commit Git commit with value-communicating message
ce-commit-push-pr Commit + push + open PR with adaptive description
ce-clean-gone-branches Clean local branches whose remote is gone
ce-worktree Manage git worktrees for parallel development

Research & Context

Skill Purpose
ce-sessions Query session history across Claude Code / Codex / Cursor
ce-slack-research Search Slack for organizational context
ce-riffrec-feedback-analysis Convert recordings/audio into structured feedback

Workflow Utilities

Skill Purpose
ce-demo-reel Capture GIF/terminal/screenshot demo reels for PRs
ce-report-bug Report a bug in the CE plugin
ce-resolve-pr-feedback Resolve PR review feedback in parallel
ce-test-browser Browser tests on PR-affected pages
ce-test-xcode Build/test iOS apps on simulator via XcodeBuildMCP
ce-setup Diagnose environment, install missing tools, bootstrap config
ce-update Check CE version, fix stale cache
ce-release-notes Summarize CE plugin releases

Development Frameworks

Skill Purpose
ce-agent-native-architecture Build AI agents using prompt-native architecture
ce-dhh-rails-style Write Ruby/Rails in DHH's 37signals style
ce-frontend-design Create production-grade frontend interfaces

Content & Tools

Skill Purpose
ce-proof Create/edit/share via Proof collaborative editor
ce-gemini-imagegen Generate/edit images via Google Gemini API

Beta/Experimental

Skill Purpose
ce-polish-beta Human-in-the-loop polish phase post-review
ce-dogfood-beta Diff-scoped browser QA: test matrix → drive app → auto-fix → commit
lfg Full autonomous engineering workflow

Agents (50+ specialized subagents, invoked by skills)

Review Agents (20)

ce-adversarial-reviewer, ce-agent-native-reviewer, ce-api-contract-reviewer, ce-architecture-strategist, ce-code-simplicity-reviewer, ce-correctness-reviewer, ce-data-integrity-guardian, ce-data-migration-reviewer, ce-deployment-verification-agent, ce-julik-frontend-races-reviewer, ce-maintainability-reviewer, ce-pattern-recognition-specialist, ce-performance-oracle, ce-performance-reviewer, ce-reliability-reviewer, ce-security-reviewer, ce-security-sentinel, ce-swift-ios-reviewer, ce-testing-reviewer, ce-project-standards-reviewer

Document Review Agents (7)

ce-coherence-reviewer, ce-design-lens-reviewer, ce-feasibility-reviewer, ce-product-lens-reviewer, ce-scope-guardian-reviewer, ce-security-lens-reviewer, ce-adversarial-document-reviewer

Research Agents (9)

ce-best-practices-researcher, ce-framework-docs-researcher, ce-git-history-analyzer, ce-issue-intelligence-analyst, ce-learnings-researcher, ce-repo-research-analyst, ce-session-historian, ce-slack-researcher, ce-web-researcher

Design Agents (3)

ce-design-implementation-reviewer, ce-design-iterator, ce-figma-design-sync

Workflow/Docs Agents (3)

ce-pr-comment-resolver, ce-spec-flow-analyzer, ce-ankane-readme-writer

Commands (1 global)

  • triage-prs.md — Triage open pull requests (.claude/commands/)

Scripts

  • src/index.ts — Bun CLI: install, cleanup subcommands
05

Prompts

Compound Engineering — Prompts

Verbatim Excerpt 1: ce-plan SKILL.md (Planning Role Separation)

---
name: ce-plan
description: "Create structured plans for multi-step tasks -- software features, research workflows, events, study plans, or any goal that benefits from breakdown..."
---

# Create Technical Plan

**Note: The current year is 2026.** Use this when dating plans and searching for recent documentation.

`ce-brainstorm` defines **WHAT** to build. `ce-plan` defines **HOW** to build it. `ce-work` executes the plan. A prior brainstorm is useful context but never required — `ce-plan` works from any input: a requirements doc, a bug report, a feature idea, or a rough description.

**When directly invoked, always plan.** Never classify a direct invocation as "not a planning task" and abandon the workflow. If the input is unclear, ask clarifying questions or use the planning bootstrap (Phase 0.4) to establish enough context — but always stay in the planning workflow.

This workflow produces a durable implementation plan. It does **not** implement code, run tests, or learn from execution-time results. If the answer depends on changing code and seeing what happens, that belongs in `ce-work`, not here.

Prompting technique: Strict role separation with explicit prohibition against scope bleed. Uses "when to trigger" heuristics + "when NOT to trigger" anti-pattern enforcement. Phase naming (Phase 0.4) creates a structured decomposition tree.


Verbatim Excerpt 2: ce-code-review SKILL.md (Multi-Mode Orchestration)

---
name: ce-code-review
description: "Structured code review using tiered persona agents, confidence-gated findings, and a merge/dedup pipeline..."
---

## Severity Scale

## Mode Detection

| Mode | When | Behavior |
|------|------|----------|
| **Interactive** (default) | No mode token present | Review, apply safe_auto fixes automatically, present findings, ask for policy decisions on gated/manual findings... |
| **Autofix** | `mode:autofix` in arguments | No user interaction. Review, apply only policy-allowed `safe_auto` fixes, re-review in bounded rounds... |
| **Report-only** | `mode:report-only` in arguments | Strictly read-only. Review and report only, then stop with no edits... |
| **Headless** | `mode:headless` in arguments | Programmatic mode for skill-to-skill invocation. Apply `safe_auto` fixes silently (single pass), return all other findings as structured text output... |

Prompting technique: Argument-token parsing pattern (mode:autofix, mode:headless) for programmatic caller detection. Headless mode explicitly designed for skill-to-skill orchestration — the output is structured text, not prose. Confidence-gated finding classification (safe_auto, gated_auto, manual, human, release) creates a tiered action routing protocol.


Verbatim Excerpt 3: ce-work SKILL.md (Plan-to-Execution Handoff)

## Phase 1: Quick Start

1. **Read Plan and Clarify** _(skip if arriving from Phase 0 with a bare prompt)_
   - Read the work document completely
   - Treat the plan as a decision artifact, not an execution script
   - Check for a `Scope Boundaries` section — these are explicit non-goals. Refer back to them if implementation starts pulling you toward adjacent work
   - **Do not edit the plan body during execution.** The plan is a decision artifact; progress lives in git commits...
   - The only plan mutation during ce-work is the final `status: active → completed` flip at shipping...

Prompting technique: "Decision artifact not execution script" framing locks the plan from mutation during execution. Scope Boundaries as an explicit named section prevents scope creep. The status: active → completed state machine is the plan's only mutable field during work.

09

Uniqueness

Compound Engineering — Uniqueness

Differs from Seeds

Closest seed: BMAD-METHOD — both use persona-md subagents, a plan-first workflow with explicit phase separation, and skills as user-invocable entry points. The key architectural deltas: (1) CE adds a compounding memory loop (ce-compound + ce-compound-refresh) as a first-class workflow phase that persists learnings across sessions — BMAD has no equivalent; (2) CE's code review is a parallel multi-persona fan-out with structured JSON dedup vs. BMAD's sequential review step; (3) CE achieves the widest cross-runtime portability in this corpus (Claude Code, Codex, Cursor, Copilot, Factory Droid, Qwen Code) while BMAD targets Claude Code primarily; (4) CE explicitly supports cross-model review gates where a different tool (RepoPrompt, Codex, Copilot) reviews the main agent's work — BMAD does not address model routing.

Positioning

CE is the premium, institutionally-oriented plugin: high component count (38 skills + 50 agents), multi-runtime, and purpose-built for teams that value the accumulation of reusable knowledge over single-session productivity gains. It is the most "opinionated about process" framework in the batch.

Observable Failure Modes

  1. Compound note rot — Without regular ce-compound-refresh runs, COMPOUND notes become stale and introduce outdated grounding into future plans. The skill lacks automatic staleness detection.
  2. Review token cost — Spawning 10–20 reviewer agents per review cycle is expensive. Teams on tight token budgets will short-circuit to ce-code-review mode:report-only.
  3. Cross-runtime feature gap — Some skills explicitly depend on Claude Code primitives (e.g., ToolSearch for deferred tool preloading). On Codex or Cursor, these paths fall back to numbered-list output with degraded interactivity.
  4. Codex install complexity — The three-step Codex install (marketplace + bun CLI + TUI) has a high abandonment risk. The README acknowledges this and promises simplification once Codex's plugin spec matures.
  5. Headless mode dependency — Skill-to-skill orchestration relies on mode:headless argument parsing. Any caller that doesn't pass the correct mode token gets interactive behavior that blocks automation loops.
04

Workflow

Compound Engineering — Workflow

Main Development Cycle

[ce-strategy] → [ce-ideate?] → [ce-brainstorm] → [ce-plan] → [ce-work] → [ce-code-review] → [ce-compound]
                                                                                ↑                     ↓
                                                                         [ce-polish-beta]    [ce-product-pulse]

Phases + Artifacts

Phase Skill Artifact
Strategy ce-strategy STRATEGY.md (target problem, approach, persona, metrics, tracks)
Ideation (optional) ce-ideate Ranked ideation doc in docs/
Requirements ce-brainstorm Requirements doc in docs/brainstorms/
Planning ce-plan Implementation plan in docs/plans/YYYY-MM-DD-NNN-<name>-plan.md
Execution ce-work Code commits, git branch/worktree, task list
Review ce-code-review Review report, per-agent JSON artifacts in /tmp/compound-engineering/ce-code-review/<run-id>/
Learning ce-compound COMPOUND notes committed to repo
Monitoring ce-product-pulse docs/pulse-reports/<timestamp>.md

Approval Gates

Gate Where Type
Branch/worktree selection ce-work Phase 1 yes-no / choice-list
Plan clarification ce-work Phase 1 (if plan unclear) freetext-clarify
Quick vs. multi-agent review ce-code-review (interactive mode) choice-list
Gated findings resolution ce-code-review interactive yes-no per finding group
Polish phase dev server check ce-polish-beta human-required (UI verification)

Bug Fix Cycle

[ce-debug "reproduce failure, trace root cause"] → [ce-code-review] → [ce-compound]

Compounding Loop

After every feature cycle, ce-compound reads: the plan, the review findings, and any corrections made during execution. It produces a COMPOUND note (markdown file) that future calls to ce-plan, ce-brainstorm, and ce-work will include as grounding context (via ce-learnings-researcher agent). The ce-compound-refresh skill periodically audits compound notes for staleness.

06

Memory Context

Compound Engineering — Memory & Context

State Storage

Type: File-based (Markdown files committed to repo)
Persistence: Project-scoped

Artifact Files Written

Artifact Location Purpose
STRATEGY.md repo root Product anchor — problem, approach, persona, metrics
Brainstorm docs docs/brainstorms/ Requirements artifacts from ce-brainstorm
Plan docs docs/plans/YYYY-MM-DD-NNN-<name>-plan.md Structured implementation plans
Pulse reports docs/pulse-reports/<timestamp>.md Time-windowed usage/performance reports
COMPOUND notes repo (path in skill) Reusable learnings for future agents
Review artifacts /tmp/compound-engineering/ce-code-review/<run-id>/ Per-run reviewer JSON + run summary
Approved plan backup .claude/state/approved_execution_plan.json Survives context compaction

Compounding Memory

The ce-compound skill is the primary memory mechanism. After any successful feature cycle or bug fix, it creates a COMPOUND note capturing:

  • What was solved
  • Why the chosen approach was correct
  • Patterns discovered
  • Anti-patterns to avoid

Future calls to ce-learnings-researcher (spawned by ce-plan, ce-work) search these compound notes to surface relevant institutional knowledge before planning or executing.

ce-compound-refresh periodically audits compound notes for staleness (keep/update/replace/archive), preventing knowledge rot.

Session History Access

ce-sessions / ce-session-historian agent can query prior Claude Code, Codex, and Cursor session transcripts for related investigation context — a cross-session memory mechanism that works without any external database.

Context Compaction Handling

ce-work explicitly handles context compaction: the approved execution plan is written to .claude/state/approved_execution_plan.json just before ExitPlanMode so it survives context clearing. On resume, the agent reads that file instead of re-entering plan mode.

Cross-Session Handoff

Plans as durable markdown files enable cross-session handoff — a new session can pick up any plan doc and continue execution. The plan's status: active → completed transition is the completion signal.

07

Orchestration

Compound Engineering — Orchestration

Multi-Agent Pattern

Pattern: Parallel fan-out with merge/dedup
Primary use case: ce-code-review spawns 10–20 tiered reviewer agents in parallel, each returning structured JSON findings; the orchestrating skill merges and deduplicates them.
Secondary pattern: ce-doc-review, ce-simplify-code, ce-resolve-pr-feedback all follow the same parallel-spawn-then-merge pattern.

Subagent Definition Format

Format: persona-md — each agent is a markdown file under plugins/compound-engineering/agents/ with a YAML front-matter block (name, description) followed by prose persona instructions.
Spawn mechanism: Claude Code's Task tool (invoked from within skill prompts).

Multi-Model Routing

ce-code-review explicitly routes to a different model/tool for cross-model review:

  • When invoked with mode:headless or from within ce-work, it may dispatch to RepoPrompt, Codex, or Copilot as a "second opinion" reviewer
  • ce-sessions / ce-session-historian can access history across Claude Code, Codex, and Cursor sessions
  • The plugin supports BYOK via each runtime's own API key configuration

No hard-coded model assignments in config files were found; model selection is left to the host runtime.

Isolation Mechanism

Git worktreesce-worktree skill and ce-work Phase 1 Option B both use git worktree for parallel development. Multiple tasks can run in parallel in isolated worktrees without file-system conflicts.

Execution Mode

Interactive-loop — The user invokes a skill, the skill may spawn subagents, returns results, and the user decides the next step. lfg provides a more autonomous chained-execution mode. ce-dogfood-beta implements a bounded autonomous loop (drive app → auto-fix → commit → repeat until green).

Auto-Validators

ce-code-review runs: lint, type-check, unit-test, security-scan (via ce-security-reviewer, ce-correctness-reviewer, ce-testing-reviewer personas)
ce-debug enforces test-first fix: write failing test, make it pass
ce-work Phase 1 instructs reading existing tests and maintaining them

Git Automation

  • ce-commit: auto-creates commits with value-communicating messages
  • ce-commit-push-pr: commits + pushes + opens PR
  • ce-work: creates feature branches or worktrees automatically
  • ce-code-review (autofix mode): applies safe_auto fixes and commits them silently
08

Ui Cli Surface

Compound Engineering — UI, CLI & Surface

CLI Binary

Name: compound-plugin (via bun run src/index.ts)
npm package: @every-env/compound-plugin
Subcommands:

  • install compound-engineering --to codex — Install CE agents into Codex profile
  • cleanup --target codex — Remove stale CE artifacts
  • cleanup --target copilot — Remove stale Copilot artifacts

This CLI is a thin wrapper for runtime-specific installation steps that the plugin marketplace protocol does not yet automate (specifically: copying agent TOML files for Codex). It is not used for running or orchestrating the plugin in normal operation.

Local UI / Dashboard

None. The plugin has no local web dashboard or TUI. All interaction is through slash commands in the AI coding assistant's chat interface.

Observability

Review artifacts are written to /tmp/compound-engineering/ce-code-review/<run-id>/ and include:

  • Per-agent structured JSON findings
  • Run summary with applied fixes, residual work, and advisory outputs
  • Run ID referenced in skill output for callers to read

IDE Integration

Works within any supported AI coding assistant's native UI:

  • Claude Code: slash commands, Skills panel
  • Codex: TOML-registered skills, agent spawning
  • Cursor: agent chat commands
  • GitHub Copilot: VS Code extension chat
  • Factory Droid / Qwen Code: native plugin chat

Cross-Runtime Architecture

The plugin.json manifest in .claude-plugin/ is the Claude Code canonical format. Cursor, Copilot, Factory Droid, and Qwen Code read the same manifest with native translation. Codex requires the additional Bun installer step to copy .toml agent definitions until Codex's plugin spec adds custom agent support natively.

Related frameworks

same archetype · same primary tool · same memory type

claude-mem (thedotmack) ★ 78k

Background worker service captures every tool call as an observation, AI-compresses sessions, and auto-injects relevant past…

pi (badlogic/earendil) ★ 55k

A minimal, hackable, multi-provider terminal coding agent that adapts to your workflows via npm-installable TypeScript Extensions…

Agent Skills (Addy Osmani) ★ 46k

Encodes senior-engineer software development lifecycle as 23 auto-routed skills and 7 slash commands for any AI coding agent.

wshobson/agents Plugin Marketplace ★ 36k

Single Markdown source for 83 domain-specialized plugins that auto-generates idiomatic artifacts for five AI coding harnesses.

TabbyML/Tabby ★ 34k

Self-hosted AI coding assistant server (alternative to GitHub Copilot) with admin dashboard, RAG-based completions, and multi-IDE…

Qodo (PR-Agent) ★ 11k

Open-source AI PR reviewer with single-call tool architecture, PR compression for large diffs, self-reflection quality gate, and…