DeskTheory is where founder-CEOs learn to run their companies on AI leverage.

What is the Compound Engineering plugin for Claude Code?

The Compound Engineering plugin is a free add-on for Claude Code that bolts a plan, build, review, learn loop plus a library of 37 skills and 51 agents onto your terminal. It's built by Kieran Klaassen and the team at Every. The premise: every task should leave the next one easier.

I install very few plugins. Most are someone's weekend project that breaks in a month. This one I keep coming back to, because it isn't a feature, it's a way of working, and the way of working is the same one I want every part of my business to run on.

Here's the idea in one line, straight from the people who built it: "Each unit of engineering work should make subsequent units easier, not harder." Read that again with "engineering" swapped for "work." That's the whole CEO pitch. (Don't let the word "engineering" make you think it's only for coders. The loop runs the same on a financial model, a hiring plan, or a content pipeline.)

What it is (in plain English)

A plugin is a bundle of capabilities you snap into Claude Code with two lines in the terminal. This one is the official Compound Engineering plugin, and it ships a lot: 37 skills and 51 agents, plus a set of slash commands that drive a repeatable loop.

The loop is the point. It walks a task through five stages: brainstorm the requirements, plan the implementation, work through the plan, review the result, then compound the learning, and repeat with better context. Each stage has its own command. /ce-brainstorm and /ce-plan think before anything gets built. /ce-work does the building. /ce-code-review checks it. /ce-compound writes down what was learned so the next run starts smarter.

The two pieces doing the heavy lifting:

• Skills are named, reusable capabilities Claude can call on. The plugin ships 37 of them, covering planning, review, research, and shipping.
• Agents are workers Claude spins up to run a chunk of the job on their own. The plan stage fans out several at once to research your existing files and the right approach in parallel, then reports back.

The team behind it says the split of effort is "80% is in planning and review, 20% is in execution." The plugin is built to enforce that ratio, so you stop letting the agent sprint off and build the wrong thing.

Why you should care as a CEO

The compounding is the asset, not the speed. A normal AI session is amnesiac: it does the work, you close the tab, and the next session starts from zero. The compound step breaks that. Every run ends by writing what it learned into files Claude reads next time, so your terminal gets a little sharper about your business every week. That's the same memory flywheel this whole site is built on, packaged as a button.

It also fixes the thing that makes AI output untrustworthy: skipping the thinking. By forcing a plan and a review around every build, the plugin keeps the agent from confidently shipping garbage. You get a clarifying-questions step up front and a quality check at the end, on every task, without you remembering to ask for either.

And it's free and MIT-licensed, one of several open tools you can layer onto Claude Code the same way you'd layer skills you write yourself.

Where you'll see it

• In the terminal, as a set of /ce- slash commands you type to run each stage of the loop.
• In your project's docs and CLAUDE.md file, where the compound step writes the learnings it wants future sessions to read.
• In the agents it spins up to research and review in parallel, instead of you grinding through it in one slow thread.
• In any task you'd otherwise run by hand in Claude Code: a model rebuild, a research memo, a draft, not just code.
• In the broader move toward reusable skills as the real unit of AI leverage.

What you should do next

Open Claude Code and install it with two lines, /plugin marketplace add EveryInc/compound-engineering-plugin then /plugin install compound-engineering, and run your next real task through /ce-plan instead of just asking.