I gave Fable 5 five thousand extra words of instructions. It thought better and failed delivery two runs out of three.

Every time my AI missed something, I added another rule. I told it to use my sources, write in my voice, check the length, read this file first, and never make that mistake again.

Every rule fixed a real problem. And over time, I didn’t notice how bloated my harness had become.

The engine got smarter. The harness didn’t.

Here is the number that made me stop. One common writing route in my local system was loading 18,384 words before it reached the guide for the platform I was actually writing for. I never typed those words in any single sitting. I accumulated them, one correction at a time, until they became a system I couldn’t see.

If you don’t know what a harness is, that’s part of the problem. A harness is the setup wrapped around the model that you or your product can actually configure: custom instructions, project files, saved prompts, memory, skills, tools, permissions, examples, and checks. It shapes the answer before you type your next prompt. It doesn’t include every hidden system inside Anthropic or OpenAI that no user can inspect.

Most of us never decide to build one. We build it accidentally, one correction at a time. We can see the latest rule we added. We can’t see the system all those rules have become.

So I built two versions of a skill that make the setup visible. The Claude version maps what Claude can access. The Codex version maps the local instructions, skills, tools, permissions, and checks available in Codex. Each one shows what it could inspect, what stayed out of view, which parts still protect the work, and which parts deserve to be combined, delayed, tested, enforced, or retired. The downloads are called Clean My AI Harness — Claude Edition and Clean My AI Harness — Codex Edition.

Here’s what’s inside:

• A guide and two runnable skills for your own setup. Point either edition at your Claude project or Codex workspace, get one plain report of what’s shaping your AI and what’s dragging, and approve changes before anything moves.

• The result that should worry you. I gave one model roughly 5,000 extra words of good instructions. It thought harder, scored better on analysis, and failed the actual delivery two runs out of three. The compact brief passed all three.

• Why your upgraded model can feel worse than a fresh chat. The chat box hides the machine. When the model gets smarter, the accumulated rules don’t get smarter with it, and you feel the drag without being able to name it.

• Six rules for maintaining a harness you can run on your own setup, whether that setup is one Claude project or a full Codex workspace.

With the skill and the guide, you’ll be able to point either edition at your own setup and get back a map, a cleanup plan, and a receipt.