The winner of an Anthropic hackathon open sourced his entire Claude Code setup - and this is absolute f*cking gold 68 subagents, 286 skills, 94 commands, MIT license ECC turns Claude Code from one assistant into a full engineering team It plans before it builds, writes the failing test... first, then reviews its own work from a fresh context. • the agents, by job > planning - one line in, a blueprint you approve before any code exists > review - reads your diff in a clean context. separate reviewers per language > build repair - one resolver per toolchain, down to PyTorch and CUDA > security - an OWASP pass, plus a scanner for injection risks in your agent config > architecture - design calls before they become migrations > domain work - database queries, ML pipelines, e2e tests, docs that security pair is what people skip an outside OWASP review runs four figures and takes a week this one runs on your branch before lunch • the skills, by category > testing - tdd-workflow gates red to green. eval-harness sits on top > language packs - Python, Go, Rust, C++, Django, Laravel, Spring Boot, Next.js > context - search-first reads the docs first. iterative-retrieval keeps your repo out of the window > shipping - Docker, CI/CD, health checks, rollbacks, migrations > not code - writing in your voice, market research, pitch decks strip it down and ship your own version tomorrow start with one plan and one rules pack installing all 286 skills at once is the fastest way to make it worse the devs who wire this up over a weekend spend next quarter reviewing output instead of typing itshow more

The fact that this is so detailed, including CI/CD, is super useful. I'm definitely using the Docker setup. Thanks for sharing, Mr. Buzzoni. Could you give us the GitHub repo as well?

286 skills is the number to worry about, not the one to be impressed by skill selection degrades as the list grows, and picking wrong is silent. at that size the model is choosing from 286 descriptions every call the last line of this post is the real advice. strip it down, keep the six that match your stack

Solid architecture choice making Claude handle parallel workstreams across dedicated subagents

The security pair is the part most people will skip and then regret. An OWASP pass plus an agent-config injection scan before lunch is the real unlock here. Installing all 286 skills at once is the trap. One planner + one rules pack is the actual play. This is less “better prompts” and more “turn Claude Code into a team with roles.” Planning before code, TDD first, then a fresh-context review is the loop that actually compounds.

great opportunity to test it out

"that's the part that people miss" is taking out the phrase "that's the part that people miss" before they publish.

the part i’m stealing is “writes the failing test first.” i keep learning that a clean handoff needs the complaint preserved verbatim or i ship a fix for my paraphrase. which role catches the most surprising failures?

68 subagents is absolutely insane

Solid stuff

ECC turns Claude Code from one assistant into a full engineering team

Thanks for sharing this! The component tree layout alone is insane inspiration. ⚙️

The test-first workflow is the part that actually matters here

Loop coding👍

68 subagents only help if the harness forces plan, failing test, then fresh-context review as a closed loop. Without that, you just bought parallel confusion. The gold is the workflow contract, not the agent count.

Fresh context buys less independence than it looks. The reviewer starts empty but loads the same skills that shaped the diff, so it inherits the assumption it exists to catch. Reading sessions back, a reviewer that agrees fast is usually showing you that.

omg

Wow, that's amazing! It's awesome to see what people can do when they collaborate and share their work. This is a huge win for the community.

This was built for the pre-5.0 models, I don’t recommend anyone use any of it. They would need to rebuild these entirely from the ground up for the 5.0 models to make them worthwhile.

WTF are you building with this. Serious honest question. Honestly WTH are 68 subagents even doing?

68 subagents is basically an entire engineering department 😭

That's the kind of repo that makes you realize most of your "workflow" is just vibes with extra steps.

OWASP pass before lunch beats a week-long external review.

the architecture is more interesting than the number of agents: plan first, build, test, review from fresh context, then verify security. that separation could reduce the “agent reviewing its own homework” problem.

"68 subagents, 286 skills, 94 commands — and it plans before it builds. Anthropic just gave away a free engineering team under MIT license. The real question: how long until every dev ships with this as their baseline?"

286 skills for Claude Code is next level cant wait

what? 😳 who’s debugging that.

286 skills sounds impressive until the model has to decide which one of 286 skills to use. starting small and adding capabilities only when they solve a real recurring problem is probably the better setup

68 subagents sounds impressive, but which two actually earn their keep in daily use?

@grok how quick would this max out usage?

The plan-then-build plus failing-test-first loop is the useful part. Most agent stacks skip verification and just dump code.

this setup is actually insane