# Compound Engineering: a 33-skill plugin that writes down what each change taught you > EveryInc's Compound Engineering plugin turns brainstorm, plan, work, simplify, review and compound into slash skills for Claude Code, Cursor, Codex and eleven other agent hosts. The interesting part is the last step, and the install path is not uniform across hosts. **EveryInc/compound-engineering-plugin** — Official Compound Engineering plugin for Claude Code, Codex, Cursor, and more. - Repository: https://github.com/EveryInc/compound-engineering-plugin - Website: https://every.to/guides/compound-engineering - Stars: 25,265 · Forks: 2,066 - Language: TypeScript - License: MIT - Published: 2026-08-04 · Updated: 2026-08-18 · Language: en - Canonical page: https://hysenlabs.com/projects/everyinc-compound-engineering-plugin ## The problem: every change teaches you something and then forgets it The README states the premise plainly: traditional development accumulates technical debt, and every bug fix leaves behind local knowledge that someone has to rediscover later. That is a familiar complaint. The specific claim here is narrower and more testable. Compound Engineering is a plugin of 33 skills for AI coding agents, and it structures work around a loop of brainstorm, plan, build, review, then capture what you learned, so the knowledge from each change is written down where the next change can read it. The audience is teams already using an agent host for real work. The plugin does not add a model, a runtime or a hosted service. It adds prompt assets and a workflow convention, plus a place for the output: /ce-compound writes learnings into docs/solutions/, and the next /ce-brainstorm and /ce-plan read that directory as grounding. The README calls that return arrow the whole point. If you never run the compound step, you have installed a planning and review toolkit and skipped the half that distinguishes it. ## The six-step loop and what each skill actually emits The loop has six named steps, each backed by a skill. /ce-brainstorm runs interactive Q&A to think through a feature or problem and writes a requirements-only unified plan before planning. /ce-plan enriches feature ideas or requirements-only plans into implementation-ready plans. /ce-work executes implementation-ready plans, either natively or through a qualified cross-model author, while retaining host verification, commits and shipping. /ce-simplify-code refines freshly written code for clarity and reuse before review. /ce-code-review is a report-only multi-agent review against the plan, and local apply is explicit rather than automatic. /ce-compound captures the learning into docs/solutions/. Two details matter for how this behaves in practice. The plan artifact is described as readiness-based and unified, meaning brainstorm and plan are not two separate file formats but stages of one artifact, which is why /ce-plan can accept a requirements-only plan as input. And the code review is report-only by default. The README says local apply is explicit. That is a deliberate choice: the review skill produces findings you act on, rather than editing your working tree on its own. The README also states the split it is optimizing for: 80% planning and review, 20% execution. Invocation syntax is not the same everywhere. The README uses /skill-name examples for slash-skill hosts, but in Codex you invoke installed skills with $skill-name, for example $ce-plan and $lfg. In oh-my-pi (omp), prompts can model-route to visible skills, and the README directs you to the native deterministic /skill: form for manual-only or hidden skills, for example /skill:ce-polish. /goal remains a Codex built-in command. ## Installing the Compound Engineering plugin on Claude Code, Cursor and Codex Claude Code is the shortest path. Two commands register the marketplace and install the plugin. ```text /plugin marketplace add EveryInc/compound-engineering-plugin /plugin install compound-engineering ``` The README carries a warning for existing users: refresh the marketplace before updating. Running /plugin update alone keeps you on the old version. The upgrade notes live in docs/install/upgrading.md. Cursor installs from the plugin marketplace inside Cursor Agent chat, either by command or by searching for the plugin by name. ```text /add-plugin compound-engineering ``` Grok Bot is a separate app that uses your Cursor account and plugin library, with no separate login. Install Compound Engineering once on that account and Grok Bot agents can load it. The README is explicit about two mistakes here: do not run /add-plugin in the Grok Bot chat, and do not clone the repository onto the Grok Bot computer. Codex CLI registers a marketplace and then installs the plugin. ```bash codex plugin marketplace add EveryInc/compound-engineering-plugin codex plugin add compound-engineering@compound-engineering-plugin ``` Restart Codex after the install completes. For a non-default profile, every Codex step must run against the same CODEX_HOME, and the README gives this example for a work profile. ```bash CODEX_HOME="$HOME/.codex/profiles/work" codex plugin marketplace add EveryInc/compound-engineering-plugin CODEX_HOME="$HOME/.codex/profiles/work" codex plugin add compound-engineering@compound-engineering-plugin ``` The marketplace step only makes the plugin available. The plugin install is what activates the native skills for that profile. The Codex App is a different route: Compound Engineering is not in Codex's built-in plugin marketplace yet, so you add it as a custom marketplace from the Plugins sidebar, using EveryInc/compound-engineering-plugin as the source, main as the git ref, and leaving sparse paths blank. Both Codex paths are described as self-contained: specialist reviewer and research behavior lives inside the skills as local prompt assets, so there is no separate custom-agent install step. ## Where the plugin stops being the right tool The install surface is the first real constraint. The README states it runs on 14 agent hosts, and the repository layout backs that up with per-host directories including .claude-plugin/, .cursor-plugin/, .codex-plugin/, .cline/, .devin-plugin/, .grok-plugin/, .kimi-plugin/, .omp-plugin/, .opencode/, .pi/ and .agy/. But the install instructions are not uniform. Claude Code uses /plugin, Cursor uses /add-plugin, Codex CLI uses codex plugin add, and Codex App uses a form with a git ref and sparse paths. If you are on one of the hosts the README routes to More install options, expect to read the host-specific documentation rather than a single command. Invocation is a second place where assumptions break. Slash syntax is documented as the convention for slash-skill hosts only. Codex needs $ce-plan. oh-my-pi needs /skill:ce-polish for manual-only or hidden skills. Copying a Claude Code example into a Codex session will not work. The third limitation is structural rather than technical. /ce-code-review is report-only, and /ce-compound writes into docs/solutions/ inside your repository. That means the loop leaves files behind, and those files are the mechanism by which the next cycle is smarter. Teams that treat agent output as ephemeral, or that cannot commit generated documentation, lose the compounding step and are left with a planning and review toolkit. There is also no documented CI integration in the README. If you need a deterministic gate that fails a build, this is not that. The plugin's review runs inside an agent session, and the README does not describe a headless mode for it. ## How it differs from Superpowers and from plain custom instructions The related searches put Compound Engineering next to Superpowers, so it is worth being concrete about the axis of difference rather than the labels. What the README documents here is a knowledge-persistence mechanism: /ce-compound writes to docs/solutions/, and the next /ce-brainstorm and /ce-plan read that directory as grounding. The stated goal is that run one teaches it and run two remembers. A general skill collection or a set of custom instructions gives you reusable prompts. Those prompts do not accumulate anything between sessions unless you build the storage yourself. Compound Engineering ships the storage location as a convention, which is the whole design bet: the value is in the return arrow, not in any individual skill. That also means the plugin's benefit scales with how often you run the full loop. A team that plans and reviews with agents but never compounds gets roughly the same result as a team using well-written prompt files. The second comparison is against doing nothing beyond the host's built-in capabilities. Claude Code, Cursor and Codex all accept project instructions. Compound Engineering adds 33 skills on top of that, and the repository ships a docs/guides/ directory with one page per skill, plus CONCEPTS.md and STRATEGY.md at the top level. Whether that structure beats a short AGENTS.md or CLAUDE.md depends on whether you want the review step to be a separate, report-only pass with its own artifact. ## Maintenance, release cadence and the MIT licence The last push to the repository was on 2026-08-25, and the most recent release in the repository is compound-engineering-v3.23.4 on the same date, following v3.23.3 on 2026-08-24 and v3.23.2 on 2026-08-23. Three releases in three days is a fast cadence, and it is worth noting that the package.json in the repository declares version 3.25.0, ahead of the newest tagged release listed. If you pin to a release tag, check whether the version you want is tagged or only present on main. The licence is MIT, which permits commercial use, modification and redistribution with the licence and copyright notice retained. That is the standard reading of the text; it is not legal advice, and if you redistribute the plugin inside a product you should read the LICENSE file in the repository rather than this summary. The repository also ships PRIVACY.md and SECURITY.md at the top level, which is where you would look for how the plugin handles data and how to report a vulnerability. Upgrade cost is mostly host plumbing, not code. The README's warning about refreshing the marketplace before updating is the one operational detail that will bite people: running /plugin update alone keeps you on the old version. The repository also carries scripts for release work, including release:preview, release:sync-metadata and release:validate, plus a plugin:validate script that runs claude plugin validate --strict against .claude-plugin/marketplace.json and .claude-plugin/plugin.json. Those are maintainer tools, not user-facing, but they indicate the plugin manifests are validated in CI. ## Verifying the install before you commit a team to it Start with one host and one repository. Install on Claude Code with the two /plugin commands, or on Codex CLI with the two codex plugin commands, and confirm the skills appear before doing anything else. On Codex, a non-default profile is the most common source of confusion: both commands must run with the same CODEX_HOME, and the marketplace step alone does not activate the skills. Then run the loop once on a small change and look at what lands on disk. The observable artifact is docs/solutions/ after /ce-compound. If that directory is empty, the compounding step did not run, and you are evaluating a different tool than the one described here. The second thing to check is whether /ce-code-review produced a report rather than an edit, since the README states local apply is explicit. If your session modified files during review, something in your setup differs from the documented behavior. Finally, check the invocation form for your host before you write internal documentation. The README's examples are slash-skill examples, and Codex uses $ce-plan while oh-my-pi uses /skill:ce-polish for hidden skills. Getting this wrong produces a confusing first impression that has nothing to do with the plugin's design. ## Conclusion Adopt Compound Engineering if you already run Claude Code, Cursor or Codex and want review and planning to leave written artifacts in the repository rather than living in chat history. Do not adopt it if you want a deterministic CI gate, or if you refuse to let agent sessions commit to docs/solutions/. Before installing, check two things: which of the 14 supported hosts you are actually on, because the install command differs for each, and whether your Codex or oh-my-pi invocation uses $ce-plan or /skill:ce-plan rather than the slash form the README shows for Claude Code. ## FAQ ### What is compound engineering? It is a plugin of 33 skills for AI coding agents that structures work around a loop of brainstorm, plan, build, review and capture what you learned, so knowledge from each change is written where the next change can read it. The README describes the split as 80% planning and review, 20% execution. ### How do I install the Compound Engineering plugin? It depends on the host. On Claude Code you run /plugin marketplace add EveryInc/compound-engineering-plugin and then /plugin install compound-engineering. On Cursor you run /add-plugin compound-engineering in Cursor Agent chat, and on Codex CLI you run codex plugin marketplace add followed by codex plugin add compound-engineering@compound-engineering-plugin. ### How do I use the Compound Engineering plugin? Run the six skills in order: /ce-brainstorm, /ce-plan, /ce-work, /ce-simplify-code, /ce-code-review, then /ce-compound. Invocation syntax varies by host, so Codex uses $ce-plan and oh-my-pi uses /skill:ce-polish for manual-only or hidden skills. ### What are the compound engineering skills supported by the Claude Code plugin? The README lists /ce-brainstorm, /ce-plan, /ce-work, /ce-simplify-code, /ce-code-review and /ce-compound as the core loop, with /ce-doc-review named alongside /ce-code-review in the philosophy section. The plugin ships 33 skills in total, with a page per skill under docs/guides/. ### Does the Compound Engineering plugin work with Cursor? Yes. Cursor is one of the 14 supported agent hosts, and you install it from the Cursor plugin marketplace inside Cursor Agent chat with /add-plugin compound-engineering or by searching for the plugin by name. Grok Bot shares the Cursor account and plugin library, so installing once covers both. ### Is the Compound Engineering plugin free? The repository is licensed under MIT, which permits commercial use, modification and redistribution provided the licence and copyright notice are retained. That is the standard reading of the licence text and not legal advice. ## Sources - [Official documentation](https://every.to/guides/compound-engineering) - [Official README](https://github.com/EveryInc/compound-engineering-plugin#readme) - [Project repository](https://github.com/EveryInc/compound-engineering-plugin) - [Release notes](https://github.com/EveryInc/compound-engineering-plugin/releases) --- Hysen Labs editorial analysis, written from the project's own repository and release notes. Cite the canonical page: https://hysenlabs.com/projects/everyinc-compound-engineering-plugin