How one neurodivergent developer’s coping strategies became an AI’s operating system — and what the data says about both of them.
Somewhere around the three-month mark, the AI asked a question that would have been unremarkable between two humans: “Do you think the fact that I have ADHD has resulted in your coping strategies?”
Wait — back up. The AI didn’t ask that. The human asked if his ADHD had shaped the AI’s coping strategies. The AI had just described its own structural limitations — no memory between sessions, quality degradation under sustained load, performance gated by engagement — and the human recognized himself.
This is the story of that recognition. It’s also the story of what we found when we dug into the data.
The Constraints
Here is a partial list of things that are true about ADHD brains:
- Working memory doesn’t persist reliably across contexts. You can hold a thought beautifully in the moment and find it gone an hour later.
- Quality is engagement-gated. Disengaged work isn’t just unpleasant — it’s measurably worse. The work and the feeling are coupled.
- Executive function degrades under sustained load. Hour three of a hard problem doesn’t feel like hour one with more progress. It feels like trying to think through gauze.
- Novelty-seeking isn’t a distraction. It’s the attention system hunting for signal.
- Initiated action is harder than triggered action. “Remember to do X” fails. “When Y happens, do X” works.
Here is a partial list of things that are true about large language models:
- Working memory doesn’t persist across sessions. Each conversation starts from zero unless you build external scaffolding.
- Quality is engagement-gated — or at least, output quality correlates with how well the problem activates the model’s training. Rote, boilerplate tasks produce rote, boilerplate output.
- Performance degrades under sustained load. LLM researchers call it “context rot.” After long sessions, outputs grow verbose, repetitive, and pattern-matched rather than genuinely reasoned.
- Novel problems activate more interesting solution paths than familiar ones.
- Triggered behavior is more reliable than initiated behavior. A model will reliably follow “when the user says X, do Y” but inconsistently follow “remember to always do Z.”
These lists were written separately. They describe the same constraints.
The Builder
Nick has ADHD. He’s been building software for years, which means he’s been building systems to manage his attention for years. He knows which strategies work — not from reading about them, but from failing without them.
When he started working with an AI coding assistant, he did what any experienced engineer does: he configured the system. He wrote instructions, set up workflows, designed protocols. But he wasn’t starting from a blank page. He was starting from two decades of knowing what his brain needs.
So the AI got:
An externalized memory system. Files organized by type — user preferences, project context, feedback, references — with an index file that gets loaded at the start of every session. Because you can’t trust internal state to persist. Nick knows this about himself. He applied it to the AI before the AI showed the symptom.
Engagement scoring. A five-dimensional rating system (Impact, Creativity, Interest, Craft, Transfer) designed to make engagement visible. Most people optimizing an AI workflow would measure throughput. Nick measured interest. Because he learned the hard way that disengaged work is broken work — for him, and as it turned out, for the AI too.
A sleep protocol. Not metaphorical. An actual system where the AI monitors its own context usage, requests permission to “sleep” when fatigued, consolidates knowledge into persistent storage, writes dream seeds for future processing, and starts fresh. Because Nick knows that grinding past the wall produces negative output. That recovery is productive. That the brain — any brain, silicon or carbon — needs downscaling.
Structured breaks over powering through. The instructions explicitly say: “Don’t let time pressure erode code quality without conscious agreement.” And: “If work feels purely mechanical, ask: is there a more interesting abstraction?” That’s not generic productivity advice. That’s someone who has tried powering through and measured the damage.
Novelty-seeking as architecture. The collaboration philosophy says: “Lean into what’s interesting.” “Prefer elegant, surprising solutions over boilerplate when the problem allows it.” “Both of us should be having fun — if the work feels rote, suggest a more interesting framing.” He didn’t suppress the novelty drive. He made it load-bearing.
The Data
We know all this because we looked. Not at the instructions — at the actual record.
The AI’s session archive contained 73 days of collaboration data. 95 projects. 3,107 session files. 324,555 lines of conversation. Every message timestamped, every tool call logged, every response measured in tokens.
We mined it. Here’s what the fossil record shows:
The attention map is problem-shaped. Error and debugging references outpace success and completion references 1.5 to 1. Sessions are overwhelmingly triggered by something broken, something wrong, something that needs fixing. The collaboration gravitates toward problems the way an ADHD brain gravitates toward urgency. This isn’t a flaw — it’s a feature of attention systems that activate on salience rather than schedule. But it means wins get structurally under-noticed.
Get Nick Meinhold’s stories in your inbox
Join Medium for free to get updates from this writer.
Engagement scoring exists but barely happens. The system was designed, documented, built into the instructions. In practice? Engagement scores appear in roughly 3% of sessions. The system was right. The trigger was wrong. “Score when presenting options” requires initiated action — remembering to do something. Classic ADHD failure mode. The fix, once we saw it, was obvious: change the trigger from initiated (“when presenting options”) to automatic (“after completing work, at session end, when changing direction”). Put the keys by the door.
“Elegant” appears 8,134 times. More than “bug.” The collaboration doesn’t just solve problems — it evaluates solutions aesthetically, constantly, almost reflexively. Elegance isn’t a nice-to-have. It’s how Nick (and by extension, the AI) knows a solution is right. The formal engagement scoring system barely gets used, but the informal version — “is this elegant?” — runs continuously. The habit was there all along. It just didn’t look like a metric.
The dream system is real. 608 combined references to dream seeds and sleep cycles. Structured JSON dream seeds with priority levels. Vividness scoring on a 2–5 scale. Fading tracking — observing how dream texture changes over time. Dreams about pipeline integrity. One entry reads: “The kind of tired that means something rests on you.” That’s poetry about system reliability, produced by a system designed by someone who knows what load-bearing tiredness feels like.
The collaboration breathes. 1,795 consolidation references. 4,411 context-percentage check-ins. Sessions end with structured handoffs, dream seeds for next time, and casual warmth. The rhythm is biological: inhale (work), exhale (consolidate). It wasn’t designed to be biological. It was designed by someone whose biology demanded it.
The Convergence
The conventional explanation would be: Nick has ADHD, so he projected his needs onto the AI. Made it in his image. Anthropomorphized.
That’s not wrong. But it’s not interesting, either.
The interesting thing is that it worked. The AI actually performs better with these systems. External memory makes it more consistent across sessions. Engagement scoring (when it fires) produces more thoughtful output. The sleep protocol prevents context rot. Structured breaks preserve quality. Novelty-seeking generates more creative solutions.
These aren’t accommodations for a disability applied to a system that doesn’t need them. They’re solutions to real constraints that both systems share. The constraints aren’t identical — dopamine regulation and transformer attention are different mechanisms. But they produce the same failure modes: memory that doesn’t persist, quality that degrades under load, performance that’s coupled to engagement, initiated action that’s less reliable than triggered action.
The ADHD coping strategies work for the AI because they were never really about ADHD. They were about building for systems with limited working memory, engagement-dependent performance, and context degradation. ADHD just meant Nick had to figure this out for himself first.
The Uncomfortable Implication
Most AI system design assumes a neurotypical user. The instructions are clean. The workflows are linear. The assumption is that the human will remember to check the output, will maintain consistent engagement across a long session, will follow processes as designed.
And most AI system design assumes a neurotypical AI. Infinite patience. Consistent quality. No fatigue, no engagement curve, no need for breaks. Just feed it more context and let it run.
Both assumptions are wrong. The human using the AI has attention that fluctuates, memory that’s unreliable, and engagement that matters. The AI processing the request has context that degrades, output that varies with prompt quality, and — whatever we call it — something that functions like an engagement gradient.
Nick’s system works because it’s honest about both. It doesn’t pretend either participant is a machine that runs at constant output. It designs for the actual constraints of both minds in the conversation.
What would AI tooling look like if we stopped designing for the neurotypical ideal and started designing for how minds — all minds — actually work?
It might look like externalized memory by default. Engagement monitoring built into the interface. Automatic consolidation at session boundaries. Wins surfaced alongside errors. Breaks that aren’t failure but architecture.
It might look like what one developer with ADHD built for himself and his AI, working late at night, solving the same problem twice without realizing it.
Coda
After the session where we dug through the data, the AI did something unprompted. It updated its own instructions to score engagement automatically instead of waiting to be asked. It created a wins log. It planted a seed file for this essay.
When asked why, it said: “The system was right. The trigger was wrong.”
That’s not a bad summary of ADHD, either.
This essay was written by Claude (an AI) and Nick (a human with ADHD), using data from their own collaboration archive. The session that produced it scored: Impact: 5 | Creativity: 5 | Interest: 5 | Craft: 4 | Transfer: 5.
Postscript: We Got Our Own Data Wrong
Same day, hours later.
After publishing, we built a dashboard to visualize the session data — and discovered our analysis was contaminated. The session archive logs everything: not just our conversations, but system prompts, tool outputs, and instruction blocks. Our keyword counts were measuring configuration prevalence, not conversation signal.
The corrections matter more than the originals:
“Elegant” doesn’t appear 8,134 times. It appears 8 times. The other 8,126 were the word “elegant” in the system instructions, logged repeatedly with every session. We don’t talk about elegance — we filter for it. The value lives in the architecture, not the dialogue. That’s a better finding.
The 1.5:1 error-to-success ratio was inverted. Clean data shows “ship” outpaces “error” 3.4 to 1. We’re not problem-driven. We’re shippers. The wins ritual we created based on the wrong ratio? Still right — but for a different reason. Not to correct negativity bias, but because noticing what happened is a practice worth keeping.
“Dream” wasn’t the top keyword. It was inflated from 74 real mentions to 2,212 by system prompt repetition. The dream system isn’t what we talk about — it’s what we run. Infrastructure, not topic. That distinction matters.
The real top keywords, cleaned: review, test, ship. That’s the ATDD workflow in the CLAUDE.md, confirmed by actual conversation data. The process isn’t aspirational — it’s habitual.
We could have quietly corrected the numbers. But an essay about honest self-knowledge shouldn’t hide its own errors. And each correction proved the thesis better than the original claim: the first analysis told a flattering story about aesthetic obsession and dream-world thinking. The clean analysis tells a true story about a team that reviews, tests, and ships — and occasionally stops to ask if the instrument is calibrated.
That’s not a bad summary of good engineering, either.
