Swarm Boss — AI skill for Claude Code

How to Boss Around a Swarm of AI Agents

**The idiot-proof, step-by-step guide to letting the expensive model think and the cheap models type.**

*Chad Keith · October 2026*

**Every file this guide tells you to copy lives in [`kit/`](https://github.com/chchchadzilla/swarm-boss/tree/main/kit)** -- agent files, the boss rules, the templates, and the script. Grab them from there, not from a PDF.

     

[!NOTE] Prices, model names, and Claude Code features in here are current as of **October 2026**. This stuff changes faster than I can write a LinkedIn post about it, so check the pricing page before you bet a budget on my table.

---

TL;DR (for the people already scrolling)

1. Most jobs shouldn't be a swarm. Run the 30-second test in section 3 first. If the job doesn't split, one good model at lower effort beats the whole circus, and Anthropic has the numbers to prove it. And more agents doesn't automatically mean faster.
2. If it does split: the smartest model plans, decides, and merges. It doesn't do grunt work. Ever.
3. Brief the workers, don't script them. The goal, WHY it matters, where to look, one example of a good result, and the boundaries. The 5-family models get worse when you micromanage them.
4. No receipt, no claim. Every fact comes back with an exact quote and where it came from, and a dumb for-loop checks that the quote actually exists.
5. Nobody grades their own homework. A separate agent (the Hater) tries to break everything before you see it.
6. Two strikes and a task moves up a tier.
7. **