My AI harness is an external executive function

I built an AI setup to save money on tokens. It turned into the memory and focus my ADHD brain is missing.

TL;DR: A messy AI agent wastes money. A messy ADHD brain wastes effort. The same thing fixes both: external structure. I built it for the agent, then I saw I had built it for myself.

Context

An "agent harness" is the plain layer around an AI coding agent. Mine has four parts:

• a CLAUDE.md: a file of rules the agent reads every time it starts
• skills: fixed step-by-step routines for tasks I do often
• routing and delegation: rules for which model does what, and when to hand a job to a helper agent
• file-based memory: notes the agent writes to disk, so it remembers things between sessions

People build this to save money and keep quality steady. An agent with no structure wastes tokens and drifts off task. Structure makes it cheap and steady. That is the whole point of it.

Why these two things match

Here is the part that surprised me. An AI agent wastes money when it works with no structure. It forgets what it already did. It drifts off the task. It redoes work. It spends effort on the wrong thing. So you give it structure: rules, saved routines, external notes, clear priorities. Now it does not have to decide everything well in the moment. The structure decides for it.

ADHD is the same failure, in a person. Under pressure I forget, I drift, I redo, I focus on the wrong thing. The weak part is my judgment in the moment. So the structure that makes the agent cheap is the same structure that makes me work.

Saving money was only the reason I started. What I really built was executive function, on disk. That is what executive function is: you do not need good judgment in the moment, because the structure already holds it for you.

The map, part by part

CLAUDE.md is memory and a safety net. I take rules seriously and I hold most of them fine. I still back every important one with a hook or a checklist, just in case. My intention is solid. My attention is not. The file does not depend on attention.

Skills are getting started. Starting is the hard part, and it is worst when I already know how to solve the task. My brain marks it as done and refuses to sit down and actually do it. A skill is the same routine every time, so there is no first step to find. It is already written. I follow it.

Routing and delegation are choosing what matters. Picking what to spend effort on is the choice I get wrong when I am overloaded. A rule makes that choice once, calmly, ahead of time.

File-based memory is memory that survives a break. An idea I do not write down dissolves in about half an hour. For years my fix was Telegram messages to myself: voice notes, links, one line about why I saved them. It worked, sort of. Notes the system writes to disk are the same thing, except they come back to me at the right moment, already in context. And an interruption, which before AI was a permanent headache, stopped erasing where I was.

The honest part

I did not plan this as help for my ADHD. I built it to stop wasting money on an agent. But the two problems have the same shape, and that is worth saying out loud.

An agent with no structure and an ADHD brain with no structure fail the same way. They forget. They drift. They redo work they already did. The difference is that the agent gets a config file, and I got a diagnosis. So I gave myself the config file too.

What this changes

Two things.

If a rule keeps failing when you rely on willpower, it does not belong in your head. It belongs in a file, a checklist, a routine, something that holds it for you. That is the same move that makes AI agents work, and no one calls the agent lazy for needing it.

And the setup that looks like "AI slop" from the outside can be the thing that lets someone finish work at all. I wrote before about AI writing as an accessibility tool. That was one piece of it. This is the whole outside brain underneath.

This very article worked that way too: I talked in a raw stream, the system built the structure. The small apps I write for myself and the diagrams I sketch start the same way.

If your head works like mine, the question was never whether to use the help. It is which part to move outside first.