Late in August someone posted a link. Affaan Mustafa, who had won an Anthropic hackathon, had open-sourced his entire Claude Code setup: 68 agents, 286 skills across 787 files, 94 commands and 144 rule files, free under an MIT license.
I run my whole business through Claude Code, so I had Claude go through all of it and tell me what was worth stealing.
I took nine pieces and skipped the other seven hundred and something. The nine were not the ones I would have guessed.
Most of it was built for software teams
He built the kit for software teams: reviewers for Go, Rust, Java and Kotlin, build-error fixers down to CUDA, test loops, rule packs for Django, Laravel and Spring Boot.
I don't write code. Claude writes code for me, and most days it doesn't write any. It answers email, moves cards on my CRM board, drafts invoices, reads transcripts. A Rust reviewer is about as useful to me as a snowplow in Florida.
The big number on the README was his. The only way to find mine was to read the thing.
The first pass found three things I didn't have
Claude read the whole repo remotely, then read my own setup, and compared. Most of the kit either didn't apply or was something I already had in a different shape. Their memory system was my memory pins. Their inbox triage was my inbox pass.
Three things had no equivalent on my side.
The first was a safety audit of the Claude Code setup itself, the settings that run Claude rather than my business or my apps. I had never audited that. It found four live passwords sitting in plain text in a config file on my Mac. Two were copies of passwords already stored where they belong. One belonged to a tool I had shut down two weeks earlier. Fixing all four took one yes from me.
The second was an audit of my own skills. I have 87 of them. The scan flagged 44 as pointing at something dead: a script that moved, or a tool I stopped using in July. I had not checked whether the shelf still matched its labels.
The third was small: a nudge when the conversation window is 70% full, so I can tidy up at a natural break instead of losing context in the middle of a task. It took twenty minutes to build.
The second pass found the real gap
Then I went back with a shorter list, the items that looked interesting on their own: an end-to-end test runner, a security reviewer, a database reviewer, coding standards, an article-writing skill, and something called Liquid Glass.
Claude mapped my repo and came back with a fact I sort of knew and had not written down. I run six real apps: a members portal, my CRM board, a social media app, a booking site, a quoting tool for a client, and a fantasy league site. They share one stack, and Claude wrote all of them.
There was no shared standard. One app's notes said "Conventions not yet established." The three worst database mistakes we had made lived as prose in one app's notes, where Claude does not look before the next database change.
I took six more pieces, each rewritten for my one stack. Claude cut 2,300 lines of generic coding rules down to 71, loaded only when it is about to write code. A database checklist now uses our own incidents as its reasons. A checker walks the live site after a change and reports red or green with a screenshot. A security pass runs before anything that touches money or member data goes live.
Liquid Glass, the one that sounded most interesting, is Apple's iOS material for native apps. It does nothing for the web, so I left it.
Inputs over outputs, again
I keep landing on the same idea: inputs over outputs.
The output of that repo is 787 skill files. The input that mattered for me was two questions I had never asked: whether the thing running my business is itself safe, and whether the apps it builds have a standard to build against.
Both answers were no. Neither needed 787 files to fix. One needed a scanner run once and a real look at what it found. The other needed 71 lines and a pointer in six files.
The audit that showed me what was missing beat any skill in the kit. Damn near everything else was noise for my situation and useful for somebody else's.
What I'd tell you
If you use these tools and someone drops a giant kit on your feed, don't install it. Have your setup read it and answer one question: what does this have that I don't?
Take that and nothing else. Rewrite it for your stack and credit the author. Write down why you skipped the rest so you don't have to decide twice.
I went looking for skills and found two holes in my own house. Patching them took the rest of the morning.
