Anthropic Hackathon Winner Open-Sources His Claude Code Setup: 68 Subagents and 286 Skills
An Anthropic hackathon winner has released his entire Claude Code setup, ECC, under the MIT license. Instead of a single assistant, you get an "engineering department": 68 subagents, 286 skills and 94 commands for review, testing and security. The order is strict: first a plan, then a failing test, then the code, and a reviewer with fresh context checks the changes. Don't switch everything on at once: start with one plan and one set of rules.
AI-processed from X: @Mr_Buzzoni; edited by Hamidun News
Most people use Claude Code like one very fast intern: describe a task, get code back, then figure out what it has done. The Anthropic hackathon winner proposes a different approach and has published all of it: his entire working setup, called ECC. The author of the X post that announced it describes it as replacing a lone assistant with a "whole engineering department." The kit includes 68 subagents, 286 skills and 94 commands, under the MIT license.
One important clarification: this is not a new model and not a release from Anthropic itself. It is a configuration on top of Claude Code, meaning a set of ready-made roles, instructions and rules. The source contains no independent measurements of how much it speeds up work or reduces bugs. For now, this is an enthusiast's account, not a benchmark.
How the agent "department" works
The idea is to split the work the way a real team splits it. First comes the plan: you give one sentence and get back a plan that you sign off on before a single line of code is written. Then the test: a failing test first, then the fix (classic TDD). Then review: every change is read by a reviewer who did not see how the code was written. Fresh context matters here. A code's author tends to miss their own mistakes, and a model has the same problem when the whole history of its reasoning sits in its context window.
The remaining roles are organized by task:
- Review. A separate reviewer for each language looks at your diff.
- Build fixing. A dedicated "fixer" for each toolchain, including PyTorch and CUDA.
- Architecture. Catches design mistakes before they turn into migrations.
- Applied work. Database queries, ML pipelines, e2e tests, documentation.
- Delivery. Docker, CI/CD, health checks, rollbacks and migrations.
- Language packs. Python, Go, Rust, C++, Django, Laravel, Spring Boot, Next.js.
There are two techniques for managing context. Search-first makes the agent check the documentation before it writes code. Iterative-retrieval keeps the whole repository from flooding the model's window. Separately, the kit includes skills that have nothing to do with code: writing in your style, market research, pitch decks.
Security: the most underrated part
The author highlights a pair of security agents: a check against the OWASP list (common web application vulnerabilities) and a scanner that looks for injection holes in the agent's own configuration. The second is especially relevant: files with rules and instructions become an attack surface themselves if the agent reads untrusted text.
The author claims that an external OWASP audit costs a four-figure sum and takes a week, while here the check finishes on your branch "by lunch." That is his estimate, and this kind of run should not be treated as a substitute for a full audit. But as a cheap first filter before a release, the idea is reasonable.
What to do with it
The author's main advice sounds boring, which is why it is worth heeding: don't enable all 286 skills at once. In his words, that is the fastest way to make results worse. Every extra skill takes up space in the context and gives the model more reasons to get distracted. Start with one plan and one set of rules for your stack, and add the rest as needed.
The practical order is as follows:
- Fork the repository and immediately delete everything that doesn't relate to your language and toolchain.
- Keep the "plan, test, review in fresh context" chain and try it on a small task.
- Add the security agents before your first release.
- Expand the kit only where you see real, recurring errors.
There is a downside, too. Someone else's configuration bakes in someone else's habits and assumptions. It may conflict with your code styles and processes. A multi-agent approach consumes more tokens than a single request. So read what you run, especially the commands and scripts: a kit of hundreds of components can't be taken on trust, like any third-party dependency.
With those caveats in mind, the kit's value is clear. It shows how an experienced Claude Code user turns a chat assistant into a pipeline with checks. Even if you don't adopt ECC wholesale, you can borrow its structure. The author sums it up this way: whoever sets this up over a weekend will spend the next quarter mostly checking work rather than typing it out by hand.
- Nov 20, 2026Daydream Launches an AI Stylist That Buys Clothes Itself From 10,000 Brands
- Nov 20, 2026Orca Bots Now Check Your Projects Every 4 Hours on Their Own, With 24 Templates to Start From
- Nov 20, 2026Developer Releases 30 SEO Skills for Claude Code and Cursor, Plus MCP Servers
- Nov 20, 2026Support team builds its own report with AI agents in 17 hours on a $20 subscription
Key Questions Answered
How the agent "department" works?
The idea is to split the work the way a real team splits it. First comes the plan: you give one sentence and get back a plan that you sign off on before a single line of code is written. Then the test: a failing test first, then the fix (classic TDD). Then review: every change is read by a reviewer who did not see how the code was written. Fresh context matters here.
What are the key details on Security: the most underrated part?
The author highlights a pair of security agents: a check against the OWASP list (common web application vulnerabilities) and a scanner that looks for injection holes in the agent's own configuration. The second is especially relevant: files with rules and instructions become an attack surface themselves if the agent reads untrusted text.
What to do with it?
The author's main advice sounds boring, which is why it is worth heeding: don't enable all 286 skills at once. In his words, that is the fastest way to make results worse. Every extra skill takes up space in the context and gives the model more reasons to get distracted. Start with one plan and one set of rules for your stack, and add the rest as needed. The practical order is as follows: Fork the repository and immediately delete everything that doesn't relate to your language and toolchain.
Build & Ship Production AI Agents Without Boilerplate Code
Start your free 72-hour trial at Hamidun Academy (academy.hamidun.com) — interactive lessons on Claude Code & Cursor, live Tech Radar, and production agent templates.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.