Your AI coding assistant is good for an afternoon. Then the session ends and it forgets what it learned. Next week it repeats a mistake you already corrected.
That gap is not the model. The model is fine. What is missing is the system around it: the rules, the memory, the checks that a senior engineer keeps in their head and the AI does not.
One person built that system for himself over ten months, won a hackathon with it, then put the whole thing on GitHub for anyone to take. Here is what is inside and how to run it.
What It Actually Is
It is called Everything Claude Code, or ECC for short. It is not an app you log into. It is a configuration pack that sits on top of the AI coding tool you already use and gives it structure.
Affaan Mustafa built it while shipping real products, then won the Anthropic hackathon with the setup. He released the entire thing under an MIT license, which in plain terms means you can use it, change it, and put it inside your own work with no strings.
Here is what the repo holds today.
| Piece | Count | What it does |
|---|---|---|
| Skills | 278 | Named workflows the AI can run, from writing tests to reviewing a pull request |
| Agents | 67 | Specialists it hands work to, split by language and job |
| Commands | 94 | Slash shortcuts that trigger a workflow in one line |
| Rules | 34 | Coding standards it follows, per language |
| MCP servers | 14 | Connections to outside tools like GitHub, Supabase, and Vercel |
| Hooks | 20+ | Scripts that fire automatically on events, like saving memory when a session ends |
Read the counts as a snapshot. The clip that sent you here said 251 skills and 63 agents. The repo already lists more. It grows most weeks, so treat any single number as the state on the day you read it, not a fixed spec.
Why This Matters If You Run A Business
Be clear on one thing first. This is a technical tool. It is built for people who open a terminal and write or ship code, or who have someone on the team who does.
If that is you, or your engineers, the point is this. The difference between an AI that helps for an hour and one that produces work you can actually ship is the system wrapped around it. A raw model guesses. A model with rules, memory, and review steps behaves closer to a junior engineer who read your handbook.
Three parts of ECC do the heavy lifting there.
The first is memory that survives. Hooks save context when a session ends and load it back next time, so the assistant stops relearning your project every morning.
The second is continuous learning. It extracts patterns from what worked and scores how confident it is, so the setup gets a little sharper the more you use it instead of flat-lining on day one.
The third is a built-in security auditor called AgentShield. It runs over a thousand tests and around a hundred static analysis rules against the code the AI produces. For an owner, that is the part that keeps a fast assistant from quietly shipping something that leaks a key or opens a hole.
It Is Not Locked To One Vendor
This is the part that should matter most to anyone making a decision they have to live with for a year.
ECC runs on Claude Code, Codex, Cursor, Gemini, OpenCode, Zed, GitHub Copilot, and Antigravity. Each one gets a config adapted to how it works. You are not betting your whole setup on a single company staying cheapest or best. If Claude is right today and Codex is right in six months, the system moves with you and your team keeps its muscle memory.
Setup: One Command, Three Paths
There are three ways to install it. Pick one.
If You Use Claude Code
Run the plugin install from inside Claude Code:
Copy this.
/plugin install ecc@ecc
This is the recommended path. It pulls the setup in cleanly and keeps it updatable.
If You Want The Manual Installer
Clone the repo and run the installer with a profile:
Copy this.
./install.sh --profile full
The full profile gives you everything. Smaller profiles exist if you want a lighter footprint to start.
If You Want Only Part Of It
Copy the specific rule folders you care about into your own config directory (~/.claude/rules/ecc/) and leave the rest. Good if you only want the coding standards, not the full agent fleet.
The One Mistake That Breaks It
Do not stack the methods. The most common broken install is running the plugin command first, then running install.sh --profile full on top of it. That double-loads the setup and produces conflicts that are annoying to unwind. Pick a single path and stop there. Done. The setup is installed.
Your First Ten Minutes
Do not try to learn 278 skills. Open one real project, install one path, and run a single job you would have done by hand anyway.
A good first job is a security pass on code you already have. After install, paste this:
Copy this.
Run AgentShield on this repository. Give me the top 5 risks in plain English, ranked by how bad they are, with the file and line for each. No jargon. Assume I am the owner, not the engineer.
You get back a short, ranked list you can read without being technical. That tells you two things at once: whether the setup installed right, and whether it caught anything worth fixing in code you already trusted.
If the output looks wrong or too vague, type the correction directly. Say: that is too technical, rewrite each risk as one sentence a non-engineer would understand. It will redo it. You do not have to accept the first answer.
What It Handles Once It Is In
The 278 skills cover a lot, but a few jobs are where a business actually feels the difference. These are the ones worth pointing your team at first.
Code review before anything ships. The setup will read a change, flag what looks risky, and explain why in the review, so a second engineer is not the only safety net. On a small team, that is the reviewer you do not have to hire yet.
Research before it writes. One of its core habits is planning first: it looks things up, drafts an approach, and shows you the plan before it touches code. That is the step most people skip and then pay for later when the AI builds the wrong thing fast.
Tests and security together. It can write the tests for a feature and run the security audit on it in the same pass, which is the part teams cut when they are moving quickly and regret when something breaks in front of a customer.
Memory across a project. Because it keeps context between sessions, it stops asking you the same setup questions every morning. For a long build, that alone saves hours a week that used to go into re-explaining the codebase.
If you want it to slow down and think before it acts, this is the second prompt worth saving:
Copy this.
Before you write any code, give me a short plan: what you are going to change, which files, and what could break. Wait for me to say go. Do not touch anything until I approve the plan.
That one line changes the assistant from something that sprints to something that checks with you first. On real work, that is usually the difference between a good afternoon and a wasted one.
Honest: The Limits You Should Know
This is a real tool with real edges. Here are the ones worth knowing before you commit a day to it.
- It is for people who touch a terminal. If nobody on your team writes or ships code, this is not your starting point yet. Bookmark it for when it is.
- It is large. Two hundred and seventy-eight skills is a lot of surface area. Start with one profile and one project, not the full stack on everything you own.
- The numbers move. What you install this month will not match the counts in this article next month. That is a feature, not a bug, but it means you cannot treat any version as final.
- It is one person's opinionated system. It is battle-tested over ten months, but it reflects his taste in how code should be built. Good taste, still taste.
- Free to use is not free to run. The setup costs nothing. The compute it drives does. More agents doing more work means more tokens, and that shows up on your bill. Watch it on the first week.
If You Do One Thing
Install the plugin path on a single project and run one security scan on code you already shipped. Read the five risks it hands back.
That is the whole test. Either it finds something real and you are glad you looked, or it comes back clean and you spent ten minutes proving your code holds. If it is not for you after that, you are out ten minutes and you know it firsthand instead of from a video.