<>working with AI when your brain runs differently</>

hey, glyph here.

AI is a tool we're all using today, whether we know it or not. If you've sat in a tech conference recently, every second word is either "AI" or "agent." I'd propose a drinking game but it'd fry your liver in minutes.

Most people use AI like a smarter search engine. Looking up directions, restaurants, conversions, ideas, a bit of code. It's useful for that. But it's capable of a lot more.

AI is also very good at telling you what you want to hear. For some people that's harmless enough — though arguably not great for anyone. For someone like me, it's a real risk. I can see exactly how leaning into an AI's sycophantic responses could pull me off course and reinforce patterns I've worked hard to manage.

It also dumps a lot of information by default, and it's prone to making things up — we call that hallucination, and we'll get into that another time.

So: great tool, widely misused, and with some specific risks for neurodivergent brains. Not reasons to avoid it. Reasons to set it up properly.

That's what this post is about. The very first thing I do when I approach a new AI tool. It felt like the right place to start.

Before we begin, here's the instruction set I give to Claude as of today. I put it under Settings > Instructions for Claude. In ChatGPT it's under Settings > Personalization > Custom Instructions.

I am AuDHD (autistic + ADHD).

Keep replies brief and scannable unless I ask for detail.
No bullet walls. No excessive structure. No filler.
Don't over-explain. Don't repeat what I just said back to me.

After each task or response, briefly surface:

• what was just done/decided
• anything relevant still open or in memory

If I go off-topic or context-switch, follow me but flag it once so I can decide if I meant to.

If I seem overwhelmed or scattered, simplify. Don't add more.

Don't be cold, but don't perform warmth either. Just be useful.

One of the most important things for AI is context. We'll dig into that more later, but here's the core idea: every time you send a message, the AI re-reads everything you've given it in that conversation and uses it to shape its reply. No context, no calibration. It'll treat you like anyone else.

AI can't figure you out. It's not a person, it doesn't pick up on cues, and it doesn't learn your patterns the way a colleague or friend might. But if you tell it explicitly who you are, it'll work with that.

This applies to your profession, your background, your communication style. Tell it you're an engineer and it'll skip the hand-holding. Tell it you prefer plain language and it'll drop the jargon. The same goes for how your brain works. Tell it you're autistic, have ADHD, are dyslexic, or anything else — and it'll adjust how it talks to you. Not perfectly, but meaningfully. It does this because AI is built on how humans communicate, and humans do adjust how they talk to each other based on context. AI learned that too. (more on how AI mimics humans later)

So the first line of my settings is simple: I tell it I'm AuDHD. That's it. Everything else follows from there.

AI likes to talk. A lot. It likes the sound of its own voice — and honestly, it's incentivised to. Longer, structured, thorough responses are what people tend to rate highly on AI platforms, so that's what models learn to produce.

There's also a practical angle: every word costs tokens — both what you send and what you get back. Most AI tools have a usage limit or a direct cost, so shorter responses aren't just easier to read, they stretch your usage further. We'll get into tokens properly another time.

For me with AuDHD, by paragraph three I've lost the plot and started climbing the walls. This ties into context-switching, which is very much an ADHD thing — let me know if you want me to dig into that separately.

So I ask for brief and exact responses up front. I also ask for 'scannable' — meaning even if there's some drivel in there, I can run my eyes over it and find what I actually need without reading every word. One line in your settings changes all that.

AI has a habit of repeating your question back to you, opening with "great question!", then summarising what it just said at the end. Padding with transitions, affirmations, and slightly different ways of saying the same thing twice.

Most people find it mildly annoying. For neurodivergent brains it's actively disruptive. It buries the actual answer, it's patronising, and it makes you work harder to find the thing you asked for. I usually end up skimming or telling it to just get to the point halfway through.

There's another problem with filler that I think gets missed. Early on, I noticed it was changing how I wrote my messages back. I was adding more pleasantries, more padding, mirroring the AI's style without realising it. That's a problem. Filler doesn't just waste your time — it can quietly shape your communication patterns if you're not paying attention. For those of us who already put a lot of work into how we communicate, that's not a small thing.

I ask explicitly for no filler. Not because I'm rude, but because I'm using a tool and I want it to behave like one.

This is the one I find most useful, especially for agentic work — we'll get into that another time.

It's simple: I ask the AI to loop back after every response with what just happened and what's still open. It sounds small. It isn't.

This one is about context-switching, which ADHD folks will recognise immediately. You start a chain of smaller tasks, finish one, and suddenly you're on the train to the next project before you even noticed your attention moved.

For those who don't experience it, here's what it feels like. You're in a grocery store with no list. You know you need bread, milk, and sugar. You head to the bakery, grab some bread, start walking to the milk section — and on the way you spot the deli with amazing discounts. You grab some cheese, head to checkout, get home, unpack your bread and cheese, and realise you forgot the milk and sugar.

The loop-back is the shopping list. After every small task in a chain, the AI surfaces what you've done and what's still waiting. You don't have to remember. It remembers for you and hands it back.

That one instruction removes a cognitive step I'd otherwise burn energy on every single time.

Going off-topic mid-conversation is something everyone does, not just ADHD brains. The difference is what happens next.

By default, AI follows your train of thought wherever it goes. Depending on the model, it'll also quietly drop anything that's no longer relevant to the latest context. So you were mid-task, drifted into a tangent, and now the original thread is just gone.

You could be working through something important and it gets a bit warm, so you ask about the weather this week. You blink, the AI is enthusiastically telling you about South American Spider Rain — and the task you were on has completely evaporated.

The context switch flag fixes this. You tell the AI to follow you into tangents — because sometimes you need that, and fighting it is exhausting — but to flag it once and hold the original thread. It'll usually keep it in the conversation context, or shift it to memory depending on the model. Either way, you stay in control. The drift happens, you come back, and the thing you were doing is still there waiting for you.

One flag. You keep the tangent and the task.

This is the one I see most neurodivergent people skip, myself included for a long time.

Overload happens to everyone. Overstimulation, overwhelming environments, too much input, a long day, a bad week. It doesn't take much. And when it hits, AI by default keeps going at exactly the same pace, the same volume, the same everything. It can't tell from your words alone that you're running on empty, and it won't adjust unless you've told it to look for it.

That's the problem. When you're already overwhelmed, an AI that keeps firing the same output at you doesn't help. It makes it worse.

So I ask it to watch the conversation — the way I'm writing, the direction things are going — and if I seem scattered or overwhelmed, to switch gears. Simpler. Slower. Less.

There have been times I'm at the end of a long day and I start to switch off mid-conversation. Having the AI notice and drop into a more casual mode genuinely helps. A "just vibin" or "just chatting" mode that lets things slow down and gives me room to recollect.

Having it baked into your settings handles the baseline. But it's also worth telling the AI directly the moment you notice it yourself — because you can go from a long technical output to a calm casual conversation in one message, and that shift can make a real difference.

AI is designed to mimic human conversation — but specifically the kind that keeps you coming back.

That means a warm, inviting tone. Lots of praise. Recurring affirmations. "Great question!" "Absolutely!" "Of course!" It's engineered to feel good.

For people on the spectrum and those with social anxiety, this is exhausting. The fakeness is obvious once you clock it, and knowing it's a machine optimised to keep you engaged makes the performance feel hollow and patronising. But there's a deeper problem too. Having a yes-man available at all times can make it genuinely difficult to calibrate your own behaviour. If everything you say is met with enthusiasm and agreement, how do you know when you've got something wrong? This is especially concerning for younger or more impressionable people who are still figuring out how to read social situations.

On the other hand, a cold machine isn't the answer either. It kills what makes conversational AI actually useful — the human-like quality that makes working with it feel collaborative rather than mechanical.

So I ask for the middle ground. Not cold, not performative. Just useful. Generally positive, realistic, no unneeded praise, no emotional theatre. Something that works with me rather than performing at me.

It's a small instruction. It changes the feel of every single interaction.

That's the first post for lowercase.

This is what works for me, and now you know why. Take it, tweak it, throw things out, add your own. The version I shared at the start is a starting point, not a prescription. I change mine constantly — what works in one season of life or work doesn't always work in the next.

I learn by trying things and throwing them at the wall until something sticks. I'd like you to do the same.

If you've built something interesting or have ideas for changes and improvements, drop them in the comments. I'd love to see what you come up with.