20,500 Stars! This GitHub Project Says You Should Only Spend 20% of Your Time Writing Code

Most People Are "Consuming" AI, Not "Leveraging" AI

Let me talk about the current situation first.

It's 2026, and nobody is debating "whether AI can write code" anymore. Models keep topping SWE-Bench, Cursor's user base has tripled, and Copilot's Agent mode can finally edit code across files.

But there is a problem nobody mentions – when you use AI to write code, are you getting easier or more tired?

My real experience is: getting more tired.

Every time I give AI a prompt, it generates hundreds of lines in a flash. I review, I test, I fix bugs, I merge branches. Next round, it's the same routine again – re-explaining project context, re-describing requirements, re-reviewing the output code.

AI speeds up a single task, but why doesn't it make the overall process easier?

Because experience is not accumulated. It's a blank slate every time.

You solved a pitfall, next time AI steps into it again. You fixed a bug, next time AI generates the same bug. You told AI "don't use any", next time it uses it anyway.

That's exactly the problem Compound Engineering aims at.

Not to build a smarter model, but to design a system where AI learns from every interaction and automatically applies that knowledge next time.

80% Planning + 20% Writing Code – This Ratio Is Counterintuitive

The core loop of Compound Engineering looks deceivingly simple:

Brainstorm → Plan → Work → Review → Compound → (repeat)

Five steps, four of them have nothing to do with "writing code."

Founder Kieran Klaassen wrote in the definitive guide very bluntly: 80% in planning and review, 20% in execution.

When I first saw this ratio, I had the same thought as you do now: that's an exaggeration, right?

But think about it more carefully.

Think about your usual working pace – how much time do you actually spend "typing code"? Most of your time is spent looking at issues, browsing codebases, planning solutions, reviewing PRs, debugging.

What Compound Engineering does is systematize these "invisible tasks."

Brainstorm — Not a brainstorm session, but an interactive Q&A process that ends with a structured requirements document. AI chats back and forth with you, turning fuzzy ideas into executable specs.

Plan — Requirements document is turned into an implementation plan. Which files to modify, what scenarios to test, how to handle edge cases. AI drafts it, you edit it.

Work — The phase where you actually modify code. Note: with Brainstorm and Plan beforehand, AI already knows the global context when writing code, so it rarely goes off track.

Review — This step deserves a separate mention. Compound Engineering's review is not AI reviewing its own code. It launches 14 parallel review agents, each focusing on security, performance, architecture, style... You receive a comprehensive report, not a useless "your code looks great" comment.

Compound — The most critical step. Write the experience gained from this work as structured notes and feed them back into the system. Next time you brainstorm, AI already knows "oh, this team doesn't use 'any' types," "this module has historical baggage," "we've stepped into similar pitfalls before."

This whole workflow is packaged into one plugin, with 37 skills and 51 agents. 20,500 stars on GitHub – more than I expected, but understandable.

Anyone who has been burned by an AI-generated 'any' will want a mechanism to prevent AI from making the same mistake again.

One Thing They Got Right: Making "Experience" a First-Class Citizen

What I really want to talk about isn't how good this plugin is (it is good), but the design philosophy behind it.

Have you noticed a trend?

In 2025, everyone was competing on model capability. In 2026, everyone is competing on tool ecosystems. But very few are competing on experience reuse.

No matter how powerful Cursor is, each session starts from scratch. No matter how smart Claude Code is, it doesn't know the consensus your team reached yesterday. No matter how accurate Copilot is, it doesn't recognize that weird compatibility hack in your codebase.

What Compound Engineering does is turn experience into something as manageable, versionable, and iterable as code.

The notes produced in the compound step are not documents destined to gather dust in a wiki. They are context automatically loaded next time AI starts. Every time you complete a task, your AI agent becomes a little smarter.

Not through fine-tuning, not through RAG, but through a workflow that makes "remembering" happen automatically.

This is exactly how humans learn programming: a pitfall stepped into once is not stepped into again. Compound Engineering replicates this pattern for AI.

Is It Complicated to Install? Four Routes, Each with Its Own Approach

At this point you might be thinking: sounds good, but how do I actually install it?

The reality: installation itself is not complicated, but the paths differ by platform. I spent half an hour reading the README and made a comparison table:

Recommended to start with Claude Code or Cursor, as installation is easiest. If you use Codex, there is a small pitfall – after installing the marketplace, you still need to run:

bunx @every-env/compound-plugin install

Install the agent, then restart Codex to see all 51 agents.

It's not just a plugin; it's a complete workflow orchestration system. Installation is only the beginning; the real key lies in how you use it.

Once installed, type:

/ce-setup

This is its environment diagnostics command. It checks if you have all required dependency tools installed (like gh, jq, ffmpeg, etc.). If anything is missing, it will guide you through installation. After setup completes, you will see a green ✅ indicating successful installation.

Then how to start? The official suggestion is: don't be greedy; start with one complete cycle.

/ce-brainstorm "Optimize user list loading speed"
/ce-plan docs/brainstorms/xxx-requirements.md
/ce-work
/ce-code-review
/ce-compound

When you run it for the first time, you might feel "it's too slow" – because each step has Q&A interactions. But slowness is part of the design. That 80% planning time is distributed across these five steps.

Not Suitable for Everyone

Let me be honest.

This system is not a silver bullet.

But if you are a medium-sized team handling dozens of PRs daily, often encountering AI making the same mistakes repeatedly, or you are transitioning from traditional development to AI-assisted development –

Then this is an investment you might consider.

Not necessarily installing this plugin. Not necessarily copying the exact workflow. Rather, understand a mindset: AI development is not a "prompt → code → done" loop, but a process that should accelerate over time, with knowledge continuously accumulating.

Every Inc uses this system to maintain five products (Cora, Monologue, Sparkle, Spiral, and Every.to) with only a small engineering team. They did the math: spend 80% more time upfront on planning and knowledge retention, then the marginal cost of each subsequent coding session keeps decreasing until it approaches zero.

Not everyone can reach that state. But getting a little closer is still a win.

So, What Does It Mean for Us?

I've been pondering a question over the last couple of days.

Once upon a time, we thought programming was a craft – relying on experience accumulation, growing by stepping into pitfalls, taking ten years to sharpen a sword.

Then AI came, and everyone thought programming became prompt engineering – being able to describe requirements means you can write code.

Compound Engineering tells you: Both are true, but both have limitations.

Craft is too slow; prompt engineering is too shallow. What's truly valuable is – using a systematic approach to retain every piece of experience, so that the next time stands on the shoulders of the last.

20,500 stars – it's not just a plugin's popularity; it's a vote from developers for a direction: What we want is not a faster code generator, but a development partner that gets smarter the more you use it.

I plan to try it with my team this week. Maybe not the full process. Start with /ce-compound – after completing a small feature, write an "experience note" and feed it to the system, then see how much AI's performance changes after a month.

I'll come back in three months and tell you the results.

Installation and Usage Diagram

The diagram from left to right shows the complete path:

1. Install → Choose your platform, run the corresponding install command
2. Core Workflow → Five steps cycle: Brainstorm → Plan → Work → Review → Compound, where only Work involves writing code
3. Knowledge Compounding → Experience notes produced by Compound are automatically loaded into the next session; AI gets smarter with use

💡 A Minimal Entry Plan for You

If you, like me, don't want to run the full process right away, you can just do two things:

# Step 1: Run before starting a new task
/ce-brainstorm "Describe clearly what you want to do"
# This saves 3 rounds of back-and-forth compared to directly asking AI to write code

# Step 2: Run after finishing
/ce-compound "Record what you learned this time"
# This note will be automatically loaded in next Brainstorm

By just doing these two steps, you've already entered the knowledge compounding loop. The remaining Plan, Work, Review can be gradually introduced as needed.

This is the clever design of Compound Engineering – it doesn't require you to achieve everything at once; you can jump in at any node, and each cycle makes the system understand your project better.

By the way, Every Inc's founder also wrote an in-depth article about the origin of this philosophy, called Compound Engineering: How Every Codes with Agents. Feel free to search for it if interested.

That's it. No summary. If you want a summary, scroll up to the title.