The promise vs. the reality
The promise: describe the overwhelming thing, get back one clear next step, instantly. That genuinely works. The reality has a shape problem: the interface is a window you have to find and open. Research on adults with ADHD finds the specific weakness is time-based prospective memory — reliably doing the planned thing at the planned time (Altgassen et al., 2014). A tool that requires you to remember it exists is standing on the very function it was supposed to replace.
Three prompts that actually help
These are the patterns worth keeping from the prompt-guide genre. All three work because they constrain the output format — an unconstrained AI answer is just a different kind of overwhelm.
- Micro-step architect. "I need to [task]. Do not give me a plan. Give me the first five physical steps, each under two minutes, checklist format, no explanation." Bypasses initiation by making the first action trivial.
- Signal-to-noise filter. "I am about to brain-dump. When I finish, extract hard deadlines, list today's actions, and park everything else. Under 100 words." Turns AI output from another wall of text into a usable filter.
- Textual body double. "Act as a silent body double for 30 minutes. I am working on [task]. Every 10 minutes ask only: what is the very next click?" This approximates the social presence that interview research shows adults with ADHD actually rely on (arXiv, 2026).
Why "better prompts" hit a wall
The prompt fixes the response, not the return. On day one the tab is open and the thread is warm. On day four you are in a meeting, the task slips, and the thread sits unvisited — the loop reopens exactly where lists fail: at the follow-up. The 2026 study of AI-augmented task scaffolding found tools built around neurotypical assumptions — stable attention, linear time, solitary self-regulation — systematically miss the relational, socially scaffolded reality of ADHD task management (arXiv, 2026).
A perfect prompt still leaves you with a perfect prompt you have to remember to run.
What the next layer looks like
The content layer — decomposition, drafting, tone-fixing — is genuinely useful and we use it ourselves. The missing layer is the cue: something that resurface the loop at the right moment, without you remembering to ask. That is implementation-intention thinking turned into a product: the trigger lives outside your brain, and the follow-up does not depend on you (evidence for if-then encoding in ADHD, medRxiv 2026).
Practical recommendation: keep ChatGPT for thinking, drafting, and decomposition — pair it with the friction scorecard context if you want the tool-level picture. The question to ask of any AI layer is not "can it answer?" but "will it come find me?"