The intersection of cognitive behavioral therapy and large language models might be the key to solving ADHD task paralysis. The intersection of cognitive behavioral therapy and large language models might be the key to solving ADHD task paralysis. I have ADHD. And the hardest part of my day isn't solving complex problems — it's starting simple ones. Sending a three-sentence email. Loading the dishwasher. Opening the document I've been avoiding for three days. These aren't hard tasks. But for my brain, each one comes with an invisible wall of accumulated dread, perfectionism, and past failures that makes initiation feel impossible. This experience has a clinical name: task initiation failure. It affects roughly 67% of adults with ADHD, and it's not a willpower problem — it's a dopamine problem. task initiation failure 67% of adults with ADHD I spent two years trying every productivity tool on the market. None of them worked. So I built one that did. The Problem With Existing Productivity Tools Most productivity software makes a core assumption: the user can look at a task and start working on it. The app's job is to organize, schedule, or remind. For ADHD brains, this assumption is fatally flawed. Here's what actually happens: You open Notion/Todoist/Asana
You see your task list
Your brain processes every task simultaneously (ADHD brains don't queue — they broadcast)
Working memory overflows
Prefrontal cortex shuts down under stress
You close the app and doom-scroll for 45 minutes You open Notion/Todoist/Asana You see your task list Your brain processes every task simultaneously (ADHD brains don't queue — they broadcast) Working memory overflows Prefrontal cortex shuts down under stress You close the app and doom-scroll for 45 minutes The tools aren't broken. They're built for a brain architecture that doesn't match ours. What CBT Taught Me About Task Design Cognitive Behavioral Therapy has a principle called behavioral activation: the idea that action precedes motivation, not the other way around. behavioral activation You don't wait until you feel like doing something. You do the smallest possible version of it, and the dopamine generated by that micro-action lowers the threshold for the next one. CBT therapists use this with depressed and ADHD patients every day. But it requires a therapist to sit with you and manually decompose your tasks into steps small enough to bypass the emotional resistance. What if an AI could do that decomposition? Designing the Prompt Architecture This is where it gets technically interesting. You can't just ask GPT to "break down a task." Generic decomposition produces steps like: Task: Write quarterly report
1. Research data
2. Create outline
3. Write first draft
4. Review and edit
5. Submit Task: Write quarterly report
1. Research data
2. Create outline
3. Write first draft
4. Review and edit
5. Submit This is useless for ADHD. "Research data" is still an overwhelming, open-ended task. An ADHD brain looks at this list and freezes just as hard as it did at "write quarterly report." The CBT-Informed Constraints I engineered the prompt system around four constraints derived from CBT and ADHD neuroscience: Constraint 1: Maximum 2-minute steps. Every step must be completable in under 2 minutes. This isn't arbitrary — it maps to the minimum viable dopamine hit. If a step takes longer than 2 minutes, the ADHD brain has time to disengage and the wall rebuilds. Constraint 1: Maximum 2-minute steps. Constraint 2: Concrete physicality. Steps must describe observable physical actions, not cognitive processes. "Think about what data you need" → fail. "Open the Q3 spreadsheet and look at row 1" → works. CBT calls this "behavioral specificity." Constraint 2: Concrete physicality. Constraint 3: Ascending activation energy. The first step must require almost zero effort — what I call a "friction-free entry point." Each subsequent step can require slightly more effort, riding the momentum of the previous micro-completion. This mirrors the CBT concept of "graded task assignment." Constraint 3: Ascending activation energy. Constraint 4: Zero perfectionism triggers. No step should imply a quality threshold. "Write a good introduction" → fail (triggers perfectionism paralysis). "Type any sentence related to the topic" → works. The permission to be bad is therapeutically critical. Constraint 4: Zero perfectionism triggers. What the Output Looks Like With these constraints, the same task becomes: Task: Write quarterly report

1. Open Google Docs and create a new blank document
2. Type "Q3 Report" at the top
3. Open the Q3 spreadsheet in another tab
4. Copy the revenue number from cell B12
5. Paste it into your doc and write "Revenue: [number]"
6. Look at the number. Type one sentence about whether it went up or down.
7. Save the doc. You now have a draft started. Task: Write quarterly report

1. Open Google Docs and create a new blank document
2. Type "Q3 Report" at the top
3. Open the Q3 spreadsheet in another tab
4. Copy the revenue number from cell B12
5. Paste it into your doc and write "Revenue: [number]"
6. Look at the number. Type one sentence about whether it went up or down.
7. Save the doc. You now have a draft started. Step 1 has near-zero activation energy. By step 4, you're already in flow. By step 7, the Wall of Awful has been breached and most users continue working on their own. The UX Decisions That Matter Most Building the AI decomposition was only half the challenge. The other half was building a UI that doesn't trigger the same paralysis it's trying to solve. One step at a time. The app shows you only the current step. Not the full list. This is counterintuitive — most apps pride themselves on giving you the "full picture." But for ADHD, the full picture is the enemy. Each visible step is a cognitive item competing for working memory bandwidth. only the current step Zero configuration. No onboarding. No preferences. No "set up your workspace." You type your task, you get steps, you start. Every setup screen is a decision point, and every decision drains from the dopamine pool you need for the actual task. No streaks, no guilt. Most habit apps use gamification — streaks, points, leaderboards. For ADHD users, a broken streak doesn't motivate a restart. It triggers shame, which triggers avoidance, which triggers more shame. We have zero tracking. You can come back after a month and there's no red badge of failure waiting for you. Results and What I Learned The tool is called Thawly (as in "thawing" a frozen brain), and here's what surprised me: Thawly Users don't need many steps. Most people only need 2-3 micro-steps before their brain's own momentum takes over. The AI generates 5-7 steps, but the majority of users break out of paralysis by step 3. Users don't need many steps. The hardest tasks aren't the biggest. Users don't typically input "write a book." They input "reply to my mom's text" or "open that envelope." The emotional weight of small tasks is what causes paralysis, not their complexity. The hardest tasks aren't the biggest. Vague input works better than specific input. "Deal with my life" generates surprisingly useful micro-steps because the AI interprets the emotional state behind the vagueness and starts with the lowest-friction physical action. Vague input works better than specific input. The Bigger Opportunity I think we're at the beginning of something important: AI as a cognitive accessibility tool, not just a productivity multiplier. AI as a cognitive accessibility tool The current wave of AI products assumes a neurotypical user. Copilot, ChatGPT, Notion AI — they all amplify existing executive function. They help you do more of what you can already do. But what about the people who can't start at all? There's an enormous underserved market of neurodivergent adults who need AI to bridge the gap between intention and action. Not to write for them, not to organize for them — but to decompose the overwhelming into the manageable, one 2-minute step at a time. If you're building in this space, I'd love to talk. The intersection of AI, clinical psychology, and accessible design is wide open. Sean Z. is the founder of Thawly, an AI-powered micro-step engine for ADHD task paralysis. Sean Z. is the founder of Thawly, an AI-powered micro-step engine for ADHD task paralysis. Thawly

This post provides insights into new product. 

Google

Asana

Why I Used CBT Principles to Design an AI That Breaks Tasks Into Micro-Steps

About Author

Comments

TOPICS

THIS ARTICLE WAS FEATURED IN

Related Stories

Behavioral Data Collection: How Gaming Might Help Study ADHD

Cerebral Seduction: How To Stop Flirting With New Ideas

Part 1: Social Media, Depression, and ADHD

How To Hack Your Brain To Get Motivated, Even When You’re Bored

Notifications: Continuous Distraction

Part 2: Cognitive Overstimulation, Depression, and ADHD

Behavioral Data Collection: How Gaming Might Help Study ADHD

Cerebral Seduction: How To Stop Flirting With New Ideas

Part 1: Social Media, Depression, and ADHD

How To Hack Your Brain To Get Motivated, Even When You’re Bored

Notifications: Continuous Distraction

Part 2: Cognitive Overstimulation, Depression, and ADHD

Light-Mode

Classic

Newspaper

Minty

Dark-Mode

Neon Noir

Minty

HN StartUps