ADHD and AI coding agents: what helps, and how to stay out of the hyperfocus trap
Developers with ADHD often say coding agents like Claude Code and Codex help with the hardest parts of the work: starting a task, holding the context in your head, and getting through the boring steps. The same agents make it easy to lose a whole day. Every step ends in a short wait, every reply is a small reward, and there’s always one more prompt. What helps: let the agent hold the context for you, give it whole tasks instead of watching each step, time-box your sessions with real breaks, keep alerts quiet, and decide in advance when you stop. This is a guide to working with the tools, not medical advice.
- What agents help with
- The hyperfocus trap
- What the research says
- A setup that keeps the help
- The prompt that remembers for you
- Questions
What agents help with
Read what developers with ADHD write about working with agents, and the same four things come up:
- Starting. A blank file is the hardest part of a task. With an agent, the first step is a sentence: “read the code and tell me how this works.” A first draft or a plan turns a vague task into something to react to, and reacting is easier than starting.
- Holding the context. You don’t have to keep the whole codebase in your head. The agent reads the files again every time, and a project file like CLAUDE.md or AGENTS.md remembers the rules you’d otherwise forget. Zachary Proser, an engineer diagnosed with ADHD at 39, calls agents his “prosthetic executive function”: instruction files hold the context, and handoff notes keep his place when he switches.
- The boring steps. Tests, renames, migrations, boilerplate: the steps that sit on a to-do list for weeks get done in minutes.
- Company without judgment. Many people with ADHD work better with someone next to them, a practice called body doubling. In a small University of Michigan study, engineers with ADHD described AI tools as a partner for long stretches of flow, without the worry about being judged or watched that comes with a human partner.
The hyperfocus trap
The same features cut the other way. An agent works in steps, and each step ends with a wait: half a minute, two minutes, five. That’s too short to do anything else properly and long enough to open another tab. Jeff Putz, who writes about coding with AI and ADHD, names these waits as the main downside: they keep pulling his attention away. Then the reply comes, it’s a small win, and the next prompt is one sentence away.
- Checking instead of working. You watch the agent think, switch away, switch back to see if it’s done. Each check is a small context switch, and they add up.
- More sessions, more pulls. Two or three sessions in parallel feel productive, but each one interrupts you on its own schedule.
- Hours disappear. The loop of prompt, wait, reply is easy to stay in long past the point where you’re thinking clearly. You look up and it’s 2 a.m., or you’ve hit your weekly usage limit by Wednesday.
- Avoidance that looks like work. A long agent session can feel like progress while the hard thing, a decision or a conversation, waits. Several developers who write about ADHD and agents warn about exactly this.
None of this is unique to ADHD; anyone can get stuck in this loop. What ADHD adds is a stronger pull into it and a harder time getting out, which is why the setup below is mostly about getting out.
What the research says
Not much yet: one small study, a few surveys, and many personal accounts.
- How common ADHD is. In late 2023, 6% of US adults, about 15.5 million people, had a current ADHD diagnosis, and more than half of them (55.9%) were diagnosed as adults, according to the CDC. Among developers, 10.6% of those who answered Stack Overflow’s 2022 survey question on neurodiversity said they had a concentration or memory disorder such as ADHD. Only about a fifth of respondents answered that question, and it’s self-reported.
- The one study on agents. In “Two’s a Crowd”, researchers at the University of Michigan interviewed 14 US software engineers with ADHD about working alongside people and alongside AI (arXiv preprint, September 2026). All 14 had used AI for coding, and 10 were positive about it. They valued accountability and flow without being judged or watched, and named Claude Code and GitHub Copilot as partners for long, uninterrupted stretches of work. The downsides they described: checking the AI’s code is still on you, the AI gives no view of whether the whole direction is right, it offers no social support, and waiting for it to generate code breaks flow. It’s 14 people, not yet peer-reviewed, and it measured no outcomes, so it describes experiences rather than proving an effect.
- Hyperfocus. It isn’t one of the diagnostic criteria for ADHD. Researchers describe it as intense, absorbed attention in which you stop noticing everything outside the task, and its costs include losing track of time and neglecting everything else. In two surveys of 251 and 372 adults, people with more ADHD symptoms reported more hyperfocus, including in front of screens. The Michigan study treats hyperfocus as a strength; none of its participants described losing hours to AI. That risk comes from personal accounts, not from research.
- Body doubling. CHADD, the US ADHD nonprofit, describes it as working alongside someone, in person or over video, so you have someone to check in with. It also notes there’s little research behind it and it doesn’t work for everyone.
A setup that keeps the help
- Start with a plan, not code. Ask the agent to read the code and write a plan before it changes anything. The plan is your way into the task, and reading it is a natural point to stop and think. In Claude Code, that’s plan mode; my full prompt is in 3 Claude Code prompts I actually send.
- Let the agent hold the context. Keep project rules in CLAUDE.md or AGENTS.md, and before you stop, ask the agent to write down where you left off (the prompt is below). Name sessions with
/renameand come back withclaude --resume. - Give it the whole task, then step away. Fewer check-ins beat more willpower. Hand over a task that takes it ten minutes or more, not one step at a time, and do something else until it needs you. What to do with each kind of wait is in What to do while Claude Code is working.
- One session at a time, two at most. Every extra session is one more thing pulling at you. More on how many sessions is too many.
- Time-box with a timer you can see. A pomodoro round, or a longer focus block, with a break you actually take. The timer decides when you stop, so you don’t have to notice it yourself. Which rhythm fits agent work is in Pomodoro with AI coding agents.
- Make alerts come to you. Turn on notifications for when the agent finishes or needs you, and stop checking. Keep them quiet during a break. How to set them up is in Claude Code notifications.
- Decide when you stop before you start. Pick the time. When it comes, write the notes, and if there’s work left that doesn’t need you, hand it to an overnight run with clear rules (the prompt). Stopping is easier when it doesn’t mean losing the thread.
The prompt that remembers for you
Send this before you stop for the day, or before a break you’re not sure you’ll come back from on time:
Before we stop, write NOTES.md:
- What we did today, in three lines.
- What's half-done, and in which files.
- The next step, small enough to start in two minutes.
- Anything I need to decide.
Keep it short. I'll read it first thing next time.The line that matters most is the third. Next time, you don’t restart the task: you start a two-minute step, and the rest follows. Open the file before anything else, or tell the agent to read it at the start of the next session.
Where Meowdoro fits
We built Meowdoro, a focus timer around the MacBook notch for people who work with Claude Code and Codex, for exactly this loop. The timer sits where you can see it, the notch shows which agent is working and which one needs you, so you can stop checking, and during a break nothing reaches you while the agents wait. It isn’t a treatment for anything; it’s a timer that knows about your agents. Try it and tell us what’s missing.
Questions
Are AI coding agents good for people with ADHD?
Many developers with ADHD say they help with starting tasks, holding context and getting through boring steps. Many also describe losing hours to the prompt-and-wait loop. The research so far is small, so treat both as experiences, not proof.
Do AI agents make hyperfocus worse?
They can. Short waits and quick replies make it easy to keep going long past the point where you’re working well. A timer with a hard stop, fewer sessions and quieter alerts help.
How do I stop checking Claude Code every minute?
Give it bigger tasks, turn on notifications for when it finishes or needs you, and do something else in between. Fewer check-ins work better than trying harder not to look.
Can Claude Code remember where I left off?
Within a session, yes, and claude --resume brings a session back. Across days, it’s more reliable to ask it to write a short notes file before you stop and read it first next time.
Related guides
Sources: Pimenova et al., “Two’s a Crowd: Human and AI-Based Copresence for Developers with ADHD”, arXiv, 2026; CDC, ADHD diagnosis, treatment, and telehealth use in adults, MMWR, October 2024; Stack Overflow Developer Survey 2022; Ashinoff and Abu-Akel, “Hyperfocus: the forgotten frontier of attention”, Psychological Research, 2021; Hupfeld, Abagis and Shah, “Living ‘in the zone’: hyperfocus in adult ADHD”, 2019; CHADD, on body doubling; Zachary Proser, “What It Feels Like To Be a Neurodivergent Engineer”, 2026; Jeff Putz, “Coding with AI and ADHD”, 2025. All read on November 15, 2026. This guide is about working with coding tools. It isn’t medical advice; if ADHD is getting in the way of your work, talk to a clinician.