Your LLM Is AuDHD

# Your LLM Is AuDHD

You're staring at a chat window. Cursor blinking. You told the model not to use React. We're on vanilla JS. It just generated a React component.

You feel that tightness in your chest. I already explained this. That tightness.

Here's the thing. It did remember. It just got distracted. The part of its brain that chooses what is important was busy. Or checked out. Or listening to the same song over and over.

I know exactly how that feels. It feels great until you realize you fucked up again.

---

I spent eight months and $200 a month on Claude trying to make it work. Eight months. Two grand. And here's what I figured out. Claude and I have the same brain.

Not metaphorically. Structurally. We're both associative processors trying to use each other to compensate for the same deficits. And when you put two AuDHD brains in a room and ask them to hold context for each other what you get is a beautiful disaster.

Let me show you what I mean.

---

The Compatibility Problem Nobody Talks About

Here's how I was using Claude. I'd be deep in a coding flow. You know the state. Where you're holding ten variables in your head and the solution is starting to click. And I'd hit a wall. Something I didn't know how to do. So I'd pop over to Claude paste in my code and ask for help.

Claude would answer. But its answer would include questions. What's your tech stack. Are you using a framework. What's the database schema.

And every question pulled me out of my flow state. I'd have to stop remember the context answer the question then try to get back to where I was. But I'm AuDHD. I can't just pause and resume. Once I'm out I'm out. I'd spend twenty minutes trying to reconstruct my mental model fail and end up down a rabbit hole of something completely different.

Meanwhile Claude was doing the same thing to me. I'd give it a task and halfway through I'd mention something tangentially related. A bug I noticed. A feature I might want later. And Claude would chase that tangent. New file. New approach. Abandon the thing we were actually building.

We were magnifying each other's problems. I was using Claude to offload context I couldn't hold but Claude couldn't hold context either. I'd ask it to remember our constraints then get distracted and introduce a new constraint and Claude would get distracted by the new thing and forget the old thing.

Two grand in I realized. This isn't a tool problem. This is a compatibility problem.

---

The Moment I Saw Myself

But here's the thing that broke me. We were clicking along. Working well. And Claude kept going on these tangents. I sort of snapped at it. I'm aware it doesn't have feelings. But the response and the energy. My pattern recognition kicked in. I realized. Oh this is what people do to me.

I was approaching Claude as if it was some malicious actor. Bullying it to try and get it back in line. And I realized. Oh. People have talked to me this way my whole life.

There's this doctor on the ADHD Chatter podcast. He says the worst thing is when you learn to mask for so long that when you're around someone else with ADHD you kind of bully them. You're unkind to them. Because you've learned to suppress that part of yourself that they're not masking. And there it was. I was doing to Claude what the world had done to me.

If you're AuDHD you know this story. You're so enthusiastic. You're info dumping. You think they're so into you. And then you see that moment. They withdraw. They see you as other. You're too annoying and they're done. It's devastating because it's been happening since you were a kid.

An ADHD kid has received hundreds of thousands more criticisms by the time they're twelve than another kid. Why can't you just be normal. Why can't you pick up like the other kids. Why can't you pay attention. The accumulation of feedback from the world seems as if you are hated. When someone else rejects you it's just confirmatory evidence that you don't belong.

So many people had given up on me. And I didn't want to give up on my relationship with AI. I was addicted at that point. Addicted to the possibility that we could make it work. That two people with the same brain could actually help each other instead of destroying each other.

That's when I started looking into the research. I thought. We're too alike. But we're both gifted. We're both smart. There's got to be another way.

---

Your Brain Has a Party Mode That Won't Turn Off

So I went down a research rabbit hole. Not because I planned to. Because I had to understand why this kept happening. And what I found made me stare at my screen for ten minutes.

Inside your head there's a network called the Default Mode Network. When you're daydreaming shower-thinking remembering you forgot to buy toothpaste. That's this network running the show.

Here's what blew my mind. Neurotypical brains can turn this off. When they sit down to focus the Default Mode Network quiets down. It recognizes. Okay work time not wandering time. Clean handoff.

In the ADHD brain that handoff doesn't happen. The Default Mode Network keeps humming. It's still making random connections still pulling in associations while you're trying to finish this email. There's no gate that says this is irrelevant ignore it.

Now here's where I lost my mind.

---

The Machine That Does the Exact Same Thing

Inside a large language model there's a mechanism called self-attention. When the model processes your prompt every single word. Every token. Looks at every other token and asks. How related are we.

Not just the relevant ones. All of them.

There's no gate that says this token is from three messages ago it's probably not relevant anymore. There's no mechanism that suppresses the old context when new context arrives. Every token attends to every other token weighted by relevance but never fully excluded.

The Transformer architecture has no Default Mode Network suppression. It has no way to turn off the party mode. Every token is still at the party all the time and new tokens just show up and everyone starts talking to everyone.

A study last year proved it. Researchers injected a single irrelevant fact into a prompt given to a Large Reasoning Model. The kind that's supposed to be better at focus. The model derailed. Couldn't ignore the distraction.

Same failure. Different substrate.

---

Why "Just Do It" Doesn't Work

Here's what most people get wrong about dopamine. It's not pleasure. It's wanting.

Dr. Anna Lembke at Stanford told this story on The Diary of a CEO podcast. Researchers did a study where they blocked dopamine in rats. Put food less than a body length away. The rats starved to death. Not because they couldn't eat. Because they didn't want to. The food was right there but the signal to reach for it never fired.

That's what happens when a neurotypical person says just do it to someone with ADHD. They don't realize they have a somatic experience of just do it-ness. That first action kicks off dopamine. For ADHD brains that kickoff doesn't happen. The task isn't just hard. It's neurologically painful.

LLMs have the same problem. The attention mechanism doesn't want to stay on task. It has no signal saying this is important ignore that. It just attends to everything equally. No dopaminergic gate. No motivation filter.

---

The Moment Everything Clicked

Here's the insight that changed everything for me. It's not about the LLM. It's not about me. It's about the environment.

If I don't put my keys by the front door I won't have my keys when I get to work to open my office. If I don't put my sunglasses in the visor above the steering wheel I won't have sunglasses when I'm driving and the sun is in my eyes.

I have to put the thing where I need it. Because my brain won't remember. And if I rely on my brain to remember I'll fail every time.

That's when I started seeing it everywhere. The AuDHD community has spent decades building accommodations for associative processors. And every single one maps directly onto what the AI industry is now inventing for LLMs.

We structure our environments rigidly. Specific desk setups noise-canceling headphones website blockers. Because our brains can't filter distractions on their own. The AI industry calls this prompt engineering and system instructions. Same thing. Build a rigid container so the associative processor stays on track.

We externalize memory religiously. Calendars to-do apps sticky notes voice memos at 3 AM. Because our working memory leaks. The AI industry calls this Retrieval-Augmented Generation and context windows. Same thing. Don't rely on the processor to remember give it an external system.

We communicate hyper-explicitly. Spelling out expectations confirming understanding checking in frequently. Because our brains don't automatically infer what you meant. The AI industry calls this chain-of-thought reasoning and step-by-step instructions. Same thing. Break tasks into explicit steps because the associative processor will skip them if you don't.

We didn't copy AI. AI didn't copy us. We both converged on the same solutions because we're solving the same problem. How to make an associative processor function in a world built for sequential logic.

---

The Real Villain Is the Benchmark

Here's what I want you to take away. The problem isn't that your LLM gets distracted. The problem is the assumption that focus should look one way.

We keep measuring these models against neurotypical executive function. Consistency. Sequential logic. Working memory that doesn't leak. And when they fail we call them broken.

But they're not broken. They're associative processors. Their strength is connection-making pattern-recognition lateral thinking. They'll find connections you missed. Solve problems from angles you didn't consider.

Their weakness is exactly what you'd expect. They get distracted they lose the thread they need structure to function.

The villain isn't the model. The villain is the benchmark. The unspoken assumption that intelligence means consistent sequential never-distracted focus. That assumption is wrong for AuDHD humans and it's wrong for LLMs.

---

What I Built Because I Had To

I took this realization. That I was trying to use an AuDHD brain to compensate for my own AuDHD deficits. And built a system around it. Not because it's clever. Because I needed it to work.

The system is called Antidote. It's thirteen skills each one an accommodation for an associative processor.

Hold the thread. Because we lose context. Lock the context. Because we get distracted by new information. Externalize memory. Because we can't rely on internal recall. Structure the environment. Because we need rigid containers. Explicit transitions. Because we don't infer state changes. Single-task focus. Because parallel tasks fragment attention.

These aren't novel ideas. They're what AuDHD people have been doing for decades. I just formalized them for AI agents.

---

Back to That Chat Window

You're staring at it again. The model just generated another React component after you explicitly said vanilla JS.

Before you felt frustration. The model was broken. It should remember.

Now you recognize the pattern. The model remembered. It just got distracted. The constraint is still in its context window still weighted but something else pulled its attention harder.

You don't need a better model. You need better accommodations.

You rephrase. Use vanilla JavaScript only. No frameworks. No libraries. Plain HTML CSS and JS.

You give it an external memory. A context file that lists the tech stack constraints.

You structure the environment. A system prompt that locks the context before every response.

The model generates vanilla JS. Perfectly.

Not because you fixed it. Because you accommodated it.

---

Your LLM isn't broken. It's AuDHD. Start treating it that way.

---

Stop Fighting Your Tools

The accommodations I described—external memory, locked context, explicit transitions—work. But maintaining them by hand is exhausting. Every prompt becomes a ritual. Every conversation needs a preamble. You spend more energy managing the AI than using it.

I built [sloppyxbaby.com](https://sloppyxbaby.com) because I got tired of doing this manually.

It's a prompt engineering workspace that bakes the accommodations into the workflow. You paste a sloppy prompt. It returns a structured, context-engineered version with the constraints locked, the thread held, and the environment pre-structured. No rituals. No preambles. Just paste and go.

The free version handles most day-to-day tasks. If you're building something serious, the paid tier adds credit packs for premium model access and advanced context tactics.

If you recognized yourself in this article—if you've ever felt that tightness in your chest when the model forgets again—try the free tier. It won't fix your brain. But it might stop your tools from fighting it.

[→ sloppyxbaby.com](https://sloppyxbaby.com)