Compound Engineering plugin — source summary

Every Inc's official "Compound Engineering" plugin for coding agents — "AI skills and agents that make each unit of engineering work easier than the last." Delivered via Telegram, ingested 2026-05-29 from the README in raw/compound-engineering-plugin.md; refreshed 2026-07-25 against the live repo (MIT, 23.5k stars / 1.8k forks, 1,084 commits).

What it is

A skillpack of slash-command agent-skills implementing the compound-engineering methodology. The loop: brainstorm → plan → work → simplify → review → compound, each cycle's documentation feeding the next cycle's brainstorming.

• /ce-strategy (durable STRATEGY.md anchor), /ce-ideate, /ce-brainstorm (a right-sized requirements doc), /ce-plan, /ce-work (execute with worktrees), /ce-simplify-code, /ce-debug, /ce-code-review (multi-agent), /ce-compound (codify learnings so the next agent doesn't relearn them), /ce-product-pulse, /ce-explain, /ce-dogfood, and /lfg (the whole loop, autonomous).
• Stated split: "80% in planning and review, 20% in execution" — "the point is not ceremony, it's leverage."

What the 2026-07 refresh changes

• 32 skills, up from the founding handful — and still no standalone agents or hooks: specialist behaviour lives inside skills as local prompt assets. A deliberate bet that the skill is the only packaging unit it needs.
• 13 install targets — Claude Code, Cursor, Codex (app + CLI), Copilot, Kimi Code, Cline, Grok Build, Devin, Factory Droid, Qwen Code, OpenCode, Pi, Antigravity — each with its own plugin manifest directory in-repo. Portability is paid for per host, not inherited from a standard.
• Cross-model dispatch: /ce-work can hand execution to a different model while the host harness keeps verification and shipping. Planning and review stay where the loop lives; only the typing travels.
• Knowledge compounding has a filesystem address — solutions land in docs/solutions/ for reuse, and /ce-code-review is report-only (non-blocking feedback before merge, not a gate).

Why it matters here

• A concrete instance of agent-skills and of spec-driven-development (its brainstorm→plan loop is spec-first), with a multi-agent review step (agent-orchestration).

• The direct sibling of superpowers: both package a whole development methodology as composable skills rather than domain expertise. The contrast is the interesting part — superpowers has the agent self-invoke skills, this one exposes them as explicit /ce-* slash commands the human drives, and it adds a step neither of the phase-per-skill packs has: /ce-compound, a skill whose product is the next run's context (self-improving-agents by durable artifact rather than weight update).

• Its central idea — each unit of work makes the next easier; codify knowledge for reuse — is compound-engineering, i.e. douglas-engelbart's bootstrapping applied to software-with-agents, and a sibling of the llm-wiki/gbrain compounding artifact idea. Reflexively, this loop is essentially how this wiki was built (brainstorm → spec → plan → execute) and how it compounds each ingest.

• Sourcing: T1 (first-party repo), but the star count and the "official" framing are popularity, not evidence. No measurement anywhere of whether the compounding actually compounds — same gap superpowers and pm-skills have.

  Retracted 2026-09-01 (claim drain). The "anywhere" half is no longer true. SkillFlow (Zhang, Shi, Huang et al., arXiv 2604.17308, 19 April 2026) benchmarks lifelong skill discovery and evolution over 166 tasks in 20 families and measures exactly this: whether an agent that keeps and reuses what it learned gets better. The answer is selective, not general — Claude Opus 4.6 goes from 62.65% to 71.08% task success (+8.43), Kimi K2.5 gains +0.60 despite using skills in 66.87% of tasks, and Qwen-Coder-Next regresses against its own vanilla setting at 44.58%. The authors' line is the one to carry: high skill usage does not necessarily imply high utility. What remains true, and is now the narrower claim: nobody has measured this plugin, and the general result says the effect depends on the model underneath, so a plugin-level claim cannot be inherited from a benchmark-level one.

Related

compound-engineering · agent-skills · superpowers · spec-driven-development · agent-orchestration · self-improving-agents · agentskills-spec · claude-cowork