Issue #32 — Less Is More: Why Constrained AI Beats Open AI for ADHD Founders
🎯 TL;DR
Georgia Tech just built an AI that works better for neurodivergent users — and the key design decision wasn't making it smarter. It was giving users fewer choices. This principle changes how you should be building and choosing AI tools as an ADHD founder.
This week: Why the blank-slate AI prompt is the worst interface for an ADHD brain, and how to redesign your AI workflow around constraints instead of openness.
Read time: 5 minutes
🧩 The Counterintuitive Finding
Georgia Tech researchers built an AI career coach specifically for people with intellectual and developmental disabilities. Published in late March 2026, the work is drawing attention in the neurodiversity and AI communities for one specific reason.
They didn't make the AI more capable to help neurodivergent users.
They made it more constrained.
Instead of giving users an open chat interface ("tell me about your career goals"), the system presents structured binary choices: "Are you looking for a new job, or figuring out career direction?" Each answer narrows to another structured choice. The AI never asks you to free-form anything.
The result: neurodivergent users navigated the process more successfully, reported less overwhelm, and completed more of the workflow.
The design philosophy, stated explicitly in the research: reduce the complex processing demands that standard platforms assume users can handle.
🧠 Why This Is an Executive Function Problem, Not an AI Problem
The blank chat prompt — "How can I help you today?" — looks like a simple question. For an ADHD brain, it's a working memory and task initiation challenge compressed into a single moment.
To answer it well, you need to:
- Hold your current context in working memory ("what was I working on?")
- Identify what you actually need help with (requires metacognition)
- Formulate a specific enough request to get a useful response
- Do all of this in the milliseconds before your attention goes somewhere else
That's four cognitively demanding steps before you even start the task you came to do. For a brain with executive function deficits, this sequence is where momentum goes to die.
"The problem with most AI tools isn't that they're too limited. It's that they put the executive function burden on the user at exactly the moment the user has the least of it."
The Georgia Tech finding is important because it's peer-reviewed evidence for what many ADHD users have been saying informally: open-ended AI interfaces are exhausting in a way that well-scaffolded interfaces are not.
⚡ The ADHD Angle
This isn't just about job-search AI. It's a design principle that applies to every AI tool in your workflow stack.
The parallel in the productivity space is already visible: Taskade's new "Executive Function Profile" lets users self-report their working memory capacity, distraction patterns, and energy fluctuations — and then automatically adjusts the AI's default behavior. Fewer levels of nesting. Shorter task breakdowns. Different reminder tones. Less text per response.
The key innovation isn't the AI. It's that the AI stops treating all ADHD users the same.
Compare the two interface approaches:
| Approach | What the AI does | What the user must supply |
|---|---|---|
| Open-ended | "How can I help?" | Context, framing, specificity, structure |
| Constrained | Presents 2-3 choices | Just pick one |
| Scaffold-first | Builds the structure, asks you to fill it | Response, not initiation |
For a neurotypical user, open-ended AI is a feature. For an ADHD user, it's often a barrier.
🔧 How to Apply This to Your Actual AI Stack
You don't need to wait for AI companies to redesign their products. You can implement the constraint principle yourself right now.
1. Replace open-ended prompts with decision trees.
Instead of starting a conversation with ChatGPT or Claude by free-typing, build a saved prompt library structured as choices:
Today I need help with:
A) Writing/drafting
B) Thinking through a decision
C) Breaking down a project
D) Reviewing something I made
Pick one. The AI now has context before you even explain anything.
2. Use "scaffold-first" prompting.
Instead of asking the AI to help you, ask the AI to give you the structure first:
"Create a structured 5-item checklist for [task]. Leave the content blank. I'll fill it in."
You're using the AI to build the scaffold, then you populate it. This bypasses the "blank page" problem entirely.
3. Evaluate tools by their default interface, not their capabilities.
When choosing AI tools, ask: what does the tool assume I bring to the interaction? If the answer is "context, structure, and clear intention," that tool has a high ADHD tax. If the answer is "just show up and pick something," the ADHD tax is lower.
Current tools that score well on this: Taskade (after setting up the Executive Function Profile), Saner.AI (answer questions about your own content rather than initiate from scratch), Notion AI (when used with pre-built templates). Tools that score poorly: most raw LLM interfaces without configuration.
4. When you're building AI into your own product: default to constraints.
If you're a founder building a product with AI features, the Georgia Tech finding is a direct product insight: your ND users (and cognitively loaded users generally) will perform better with constrained interaction patterns than open ones. Progressive disclosure — start narrow, unlock complexity — is the right pattern.
🎯 This Week's Challenge
Audit one AI tool you use regularly.
Ask yourself: when I open this tool, what does it ask me to bring? How much executive function does the default interface require before I even start the actual task?
Then do one thing: either configure it to be more constrained (templates, saved prompts, structured workflows) or replace it with a tool that scaffolds by default.
You don't need better AI. You need AI that meets you where your executive function actually is.
See you Tuesday, L-P
P.S. — The Georgia Tech paper is available via Georgia Public Broadcasting (March 27, 2026). Taskade's Executive Function Profile is live now — it takes about 5 minutes to set up and the difference in default behavior is noticeable. If you've already built a constrained-AI workflow, I want to hear what it looks like. Reply and tell me.
Divergent — Strategy for brains that don't do boring.