Two hundred and forty thousand stars on GitHub. Two hundred and eighty-four skills, sixty-eight agents, ninety-four commands, twenty-two rule packs - one common set plus one per language. Adapters for sixteen environments - from Claude Code and Codex to Zed, Qwen and Kimi, though they are far from equal, as we will see.
The first reaction to numbers like that is disbelief. The second, once you look properly, is different: this is not "another prompt pack" but a warehouse of ready parts you are meant to assemble your own set from. And the main question to ask of it is not "what is inside" but "is it worth installing and what will it cost".
ECC does not make the model smarter and is not a separate AI coder. It is a layer of configuration and behaviour on top of the environment you already use. The author calls the project an agent harness performance optimisation system - and "optimisation" here is about the surroundings, not the model.
What it offers instead of a big prompt
A system prompt used to be enough: write tests, do not break the API, read the code first. A modern agent does not work that way - it calls tools, runs commands, reaches MCP servers, loads skills, spawns sub-agents and reacts to session events. The prompt stopped being text and became infrastructure.
ECC lays that infrastructure out in layers, each with its own job.
| Layer | What it solves | How much is in the repository |
|---|---|---|
| Skills | repeatable procedures: documentation lookup, review, languages, frameworks, operations | 284 |
| Agents | specialised performers for a particular kind of work | 68 |
| Commands | entry points for the human | 94 |
| Rules | standing constraints and project conventions |