Tech
A Thermonuclear ADHD Amplifier: The Same Tool That Scatters You Can Make You Finish
A developer this month built sixteen little apps in a weekend and shipped none of them. Each went from idea to working code in under an hour, then got a new tab and was forgotten. He gave the feeling a name that spread fast: a coding agent, he wrote, is a thermonuclear ADHD amplifier. Read the replies, though, and a stranger thing happens. People with actual ADHD show up to say the opposite — that for the first time in their lives, they're finishing things. Same tool. Opposite outcome. The interesting question isn't which camp is right. It's what, exactly, decides which camp you land in.
The cheap-reward machine
Start with why the amplifier metaphor lands. A coding agent collapses the gap between a whim and a thing that runs. You think "I wonder if I could…" and forty minutes later there's a working prototype with tests and a README. That is, in the most literal sense, a slot machine for builders: an instant payout for almost no input, and no friction telling you to stop. The reward circuitry doesn't distinguish a shipped product from a dopamine hit; it just wants the next pull. So the tool doesn't make you lazy — it makes you start, endlessly, and starting is the cheapest, most seductive part of any project. You spin up ten tabs, three of them on screens you're not even looking at, each a half-built world you have no realistic plan to maintain. The developer who coined the phrase had built sixteen of these in short order; he concluded the honest fix might be to cancel the subscription, because a tool that hands you cheap reward for almost no input, with no friction, had become a liability rather than an asset. That's the amplifier at full tilt: not a productivity engine, but a faster way to feel busy.
What the tool actually amplifies
Here is the move people miss: the agent is not adding ADHD to anyone. It is amplifying whatever attention pattern you brought to it. This is the oldest idea about tools — a medium shapes how its maker works, the way oil paint let painters labor slowly over invisible detail while fresco forced them to commit fast. Twenty years ago, in Hackers and Painters, Paul Graham argued the deep skill of programming is having "the whole program in your head" — holding the entire structure in active memory so you can see the consequences of a change before you make it. That is a focus skill, not a typing skill. The agent removes the typing. It does nothing for the holding. If you walked in able to hold one program whole, the tool hands you a faster body for the same focused mind, and you finish. If you walked in chasing novelty, it hands you nine more programs to not-hold, and you scatter. The wedge between the two camps was already inside you; the machine just turned up the gain.
Why ADHD builders sometimes win
So why do some ADHD developers report the agent as a salve rather than a curse? Because their bottleneck was never the idea — it was the dreary middle. The classic ADHD failure mode in building software is the cliff between an exciting start and a tedious slog: the boilerplate, the config, the fortieth near-identical CRUD form, the part with no novelty to feed on. That's precisely the stretch where the project used to die. The agent eats that stretch. It keeps the work continuously novel enough to stay engaging while quietly carrying the boring weight, so the person reaches "done" before boredom reaches them. The same instant-reward loop that scatters the novelty-chaser becomes a bridge across the exact gap that used to swallow the finisher. The tool didn't change. The pre-existing relationship between that person and the boring middle did. One person in the thread put it plainly: AI had become a salve for their mind, the stimulation they used to seek in a dozen abandoned tabs now spent on carrying a single project to the end. That is the inverse of the sixteen-app weekend, produced by the identical software — which is exactly why blaming or praising the tool misses the point.
Don't outsource the holding
The trap, then, isn't using the agent. It's letting it do the one part you can't afford to delegate: keeping the whole thing in your head. It's frighteningly easy to let the model hold the structure for you — to ask it what your own code does, to approve diffs in a system you no longer mentally own. That feels like flow and is actually its opposite. When you stop holding the program, you stop being able to see which change breaks what, and you've quietly traded the deep skill for the shallow one. Graham's makers sweated invisible detail because they carried the whole canvas in mind; a builder who has offloaded the canvas can only react to whatever the agent surfaces. Speed without holding isn't leverage. It's just a faster way to lose the thread — and the more capable the agent gets, the easier it becomes to never notice you've let go.
What this means for you
If you build with these tools, the discipline that pays is brutally simple and entirely under your control: defend the holding, and ration the starting. Concretely — finish one thing before you let yourself open the next; keep, at most, the number of live projects you can actually carry in your head, and treat a tenth tab as a tax, not a trophy; when the agent hands you a working draft, spend the saved time understanding it, not generating the next one. The novelty-chaser and the finisher are using the identical tool; the only difference is whether attention is spent holding or spent scattering. The good news is that this isn't a fixed trait you were dealt. Which fork you take is a choice you remake every session — and now you know exactly where the switch is.
The agent amplifies whatever attention you bring it. Bring scatter and it scatters you ten ways; bring the whole program in your head and it finally lets you finish.
What's scarce was never generation speed. It's the focus to hold one thing whole.
Source: framework from Paul Graham, Hackers and Painters; the "thermonuclear ADHD amplifier" line was written by David Wilson and amplified by Simon Willison (May–June 2026), alongside contrasting reports from ADHD developers. A reflection, not medical or psychological advice.