Gifamoss@gifamoss

Follow

ECC Claude Code setup 68 agents

The part of ECC that got me was the audit timeline. OWASP sweeps on your agent config plus injection checks, wrapped up by lunch, when hiring a firm for the same thing means waiting about a week and signing a four-figure check 🥶 It's a Claude Code configuration from the winner of Anthropic's hackathon, given out at no cost under MIT. 68 subagents, 286 skills, 94 commands, all running inside one window like a full development team. The workflow rules are strict in a way i respect: a plan has to be signed before anything gets built, a patch waits for a test to fail first, and a merge needs a stranger to have read it. The planner turns one line of input into that signable plan before any code exists. Each language gets its own reviewer reading diffs cold. Build doctors sit per toolchain, PyTorch and CUDA among them. An architect flags design flaws early, back when a fix costs minutes instead of a migration. Specialists handle SQL, ML pipelines, end-to-end suites and docs. Skill packs cover Python, Go, Rust, Django, Next.js, Docker, CI/CD and testing frameworks. Flip all 286 on at once though and the setup gets worse than what you had. So the integration notes say fork the repo at github.com/affaan-m/ECC, read the README before copying anything, and start with just the planner plus rules for your main language. Fill config files with only the tools you need over the next seven days. Then add code reviewers and security reviewers. Test in disposable repos, not production. Bloated? Trim. Missing something? Add a single piece at a time. And if an agent has no clear job, it doesn't get a slot. Master it and the role shifts from producing work to approving it within a quarter, or something else.

1h

Gifamoss@gifamoss

Follow

ECC Claude Code setup 68 agents

The part of ECC that got me was the audit timeline. OWASP sweeps on your agent config plus injection checks, wrapped up by lunch, when hiring a firm for the same thing means waiting about a week and signing a four-figure check 🥶 It's a Claude Code configuration from the winner of Anthropic's hackathon, given out at no cost under MIT. 68 subagents, 286 skills, 94 commands, all running inside one window like a full development team. The workflow rules are strict in a way i respect: a plan has to be signed before anything gets built, a patch waits for a test to fail first, and a merge needs a stranger to have read it. The planner turns one line of input into that signable plan before any code exists. Each language gets its own reviewer reading diffs cold. Build doctors sit per toolchain, PyTorch and CUDA among them. An architect flags design flaws early, back when a fix costs minutes instead of a migration. Specialists handle SQL, ML pipelines, end-to-end suites and docs. Skill packs cover Python, Go, Rust, Django, Next.js, Docker, CI/CD and testing frameworks. Flip all 286 on at once though and the setup gets worse than what you had. So the integration notes say fork the repo at github.com/affaan-m/ECC, read the README before copying anything, and start with just the planner plus rules for your main language. Fill config files with only the tools you need over the next seven days. Then add code reviewers and security reviewers. Test in disposable repos, not production. Bloated? Trim. Missing something? Add a single piece at a time. And if an agent has no clear job, it doesn't get a slot. Master it and the role shifts from producing work to approving it within a quarter, or something else.

1h