Thin Harness, Fat Skills

**Thin Harness, Fat Skills: The Real Reason Some AI Builders Are 100x Faster**

There’s a bold claim floating around from Steve Yegge, shared by Garry Tan on X. People using AI coding agents are **10x to 100x more productive** than engineers using standard chat tools today, and even *1000x more productive* than Googlers back in 2005.

That’s not hype. You can read the original post here:
https://x.com/garrytan/status/2042925773300908103?s=52

But here’s where most people get it wrong. It’s not about smarter models.

The same model. The same Claude. The same underlying intelligence.

The difference is the **architecture around it**.

Yegge calls it *thin harness, fat skills*. And once you see it, you can’t unsee it.

Let’s break that down.

The **harness** is the lightweight wrapper around the model. It runs the loop, manages context, reads and writes files, enforces guardrails. That’s it. Keep it thin. Simple. Fast.

The real magic lives in the **skills**.

A skill is not just a prompt. It’s a reusable procedure written in plain language, like a method call in software. It encodes judgment. Process. Domain expertise. You pass in parameters, and it executes the thinking pattern consistently.

Same skill. Different inputs. Completely different outcomes.

Then there are **resolvers**, which load the right context at the right time. And a ruthless separation between **latent work** (judgment, synthesis, reasoning) and **deterministic work** (SQL, math, compiled code). Put each in the wrong place and your system quietly falls apart.

One concept stood out to me: if you ask your AI to do something twice, you’ve already failed. The first time is manual. The second time becomes a skill. From then on, it’s a permanent upgrade.

That’s how productivity compounds. Not because the model got smarter overnight, but because your system did.

And here’s the exciting part… as models improve, every well-designed skill improves automatically.

Build it once. Let it run. Keep the harness thin. Grow the skills.

That’s where the 100x lives.