ADHD Prompts for Claude: 9 I Run and the Failure Each One Guards

Part of my Neurodivergent Systems series on tools and workflows for technical work.

Short answer: the useful ADHD prompts for Claude are the ones that name a specific failure and guard against it. “Be concise” guards nothing. “Show me the evidence before you say this is done” guards a failure I hit every week.

I have ADHD and I use Claude all day, for code and for everything around code. Over time the way I prompt it hardened into rules. Those rules live in a file my developer setup reads automatically, and I wrote them up in the Claude Code ADHD workflow. This page is the same rules as plain prompts, for anyone, sitting inside the ADHD engineer’s productivity system, with the failure each one is for and the situation where it stops working.

Nine prompts. Copy the ones whose failure you recognize. Skip the rest.

How to read this list

Every entry has three parts: the prompt, the failure it guards, and when it breaks. The third part is the one every other list leaves out, and it is the one that matters, because an ADHD tool that works on a good day and quietly stops working on a bad day is worse than no tool. You stop keeping the backup in your own head.

Where a prompt has a number in it, the number is mine. Change it to yours.

Starting

These two handle task initiation, the wall most people mean when they say “ADHD”.

1. The brain dump

I am overwhelmed by [the thing]. Here is everything in my head about it,
in no order:
- ...
- ...
Turn this into steps. Each step should take under one minute to start.

Guards: the sorting problem. Organizing a pile is the executive-function task ADHD makes expensive, so hand the pile over unsorted. Sorting it yourself first is the step that never gets taken.

Breaks when: the dump is too clean. If you tidy your thoughts before pasting them, you have already paid the cost the prompt was meant to remove. Paste the mess.

2. The first door

Of those steps, give me only the first one. Do not show me the rest.

Guards: the hallway problem. A twelve-step list is a new overwhelm. You wanted one door. This prompt makes Claude hide the hallway.

Breaks when: the first step is not actually a one-minute start. If Claude gives you “design the schema” as step one, push back: “smaller, what is the ten-second version of that?” The step has to be something your hands can begin before your brain objects.

Before I ran these two, breaking a project down took me thirty minutes of staring and giving up. Now I describe the outcome, get four or five pieces back, and pick one. The decision changes from “what do I do?” to “which of these four first?” The second decision is a hundred times easier.

Staying on task

These four keep the conversation from becoming the distraction.

3. One question

I have several questions. I will send them one at a time.
Answer only the question in front of you.

Guards: the blended answer. Three questions in one message get one half-answer to each, and untangling it is more executive-function work. One question, one answer, no parsing.

Breaks when: the questions genuinely depend on each other. If question two only makes sense given the answer to one, send one, read the answer, then decide whether two still exists. Half the time it does not.

4. Evidence before done

Before you tell me this is done, show me the evidence.
Quote the exact part that does what I asked.

Guards: confident wrong answers. Claude is fluent when it is right and exactly as fluent when it is wrong. For an ADHD brain, fluency feels like closure, and closure is what you are craving, so you accept it to make the open loop stop. This prompt keeps the loop open until the evidence closes it.

Breaks when: you are tired and skip it. That is the exact moment it matters. I keep this rule in the file my developer setup reads so that it applies even when I would not have typed it.

5. Explain before you change anything

Before you change anything, explain what you are about to change and why,
in three sentences. Then wait for me.

Guards: rubber-stamping. When output is confusing, the ADHD reflex is to accept it and move on, because understanding it is effortful and moving on feels like progress. This prompt makes the explanation arrive before the change, so you never have to reverse-engineer what happened.

Breaks when: you use it on every tiny step. It adds a round trip. Use it when the change is big, unclear, or hard to undo. For a rename, skip it.

6. The checkpoint

It has been 45 minutes. Summarize in three lines:
what we were doing, what is next, what is blocking.

Guards: hyperfocus and drift. A tool that can build anything is an endless supply of novelty. Forty minutes into “plan my week” you are designing a system for tracking weekly plans. The checkpoint forces a re-read of the original goal.

Breaks when: nothing fires it. Claude does not know 45 minutes have passed. You need a timer, or a rule in your setup that asks for this on a schedule. Without the trigger, this prompt is a good intention.

Why 45? That is roughly the size of a task I can hold in one piece. Work that fits in 45 minutes gets finished. Work that needs three hours gets abandoned at hour two.

Coming back

These two are the only prompts on this list that survive an interruption. The others assume you are still in the chair.

7. The re-entry note

Typed before the interruption, not after:

Pausing. Working on [X], stuck on [Y], next is [Z].
When I come back, remind me of this and give me the next step only.

Guards: the rebuild. Gloria Mark’s research at the University of California, Irvine, put the cost of getting back to a task after an interruption at about 23 minutes. For an ADHD brain that number is the floor. The note turns a 23-minute rebuild into a thirty-second re-entry.

Breaks when: you write it after the interruption. After is too late; the context is already gone. My rule: an interruption under a minute needs no note. Over a minute, the note goes down first, then the interruption gets handled.

8. What am I forgetting

Here is what I have done and what I plan to do next.
What am I probably forgetting?

Guards: silent drops. ADHD working memory loses things without telling you, because noticing the loss is the exact function that failed. A second processor with the whole context in front of it catches what you cannot.

Breaks when: Claude does not have the context. If the conversation is fresh, it is guessing from generic patterns and the answer is noise. Paste the checkpoint first, then ask.

Closing the day

9. What did I actually get done

Here is my list from this morning and here is what happened.
Tell me what I actually completed, in plain terms, before you tell me
what is left.

Guards: discounting the positive. At the end of a day my brain reports “nothing” regardless of what happened. Asking for the completed list first, and the remaining list second, is a small reorder with a large effect. In my developer setup this prompt became a skill that reads the day out of the version-control history instead of my memory, because my memory is the unreliable part. It is described in the skills I run.

Breaks when: you feed it your feelings instead of your record. “I feel like I did nothing” is the input that produces “you did nothing”. Feed it the list and the events. Let it read the record, not the mood.

How to make them stick

A prompt you have to remember to type is a prompt you will not type on the day you need it. The bad day is the day working memory is at its worst, and the prompt list is the first thing to fall out of it. So the last step is to stop typing them.

There are three levels, from cheapest to most durable.

Level one: a text file. Keep the nine in one note. Pin it. When you open Claude, paste the two or three that match the day. This costs nothing and fails only when you forget the note exists, which on a bad day is likely.

Level two: a project or custom instruction. Claude’s apps let you set standing instructions for a project or a workspace. Put the four staying-on-task prompts there as rules: one question per message, evidence before done, explain before changing, three-line checkpoint on request. Now they apply without typing. The brain dump and re-entry prompts stay manual, because they need your input in the moment.

Level three: a file your tooling reads. For developers, Claude Code reads a CLAUDE.md file at the start of every session. My version has a section headed Rules (ADHD-Friendly) with the five rules the prompts above descend from. It has a Current Checkpoint section that gets rewritten every 45 minutes, which is the re-entry note in permanent form, and a Known Learnings section so that the same mistake is not re-explained twice. That last section is the one I did not expect to matter. It turns out that half my re-explaining was to the tool, not to a person, and a file ended it.

Whichever level you pick, the test is the same: does the rule fire on a day when you would not have typed it? If the answer is no, you have a list of good intentions, and this page has plenty of those already.

What is not on this list

No “act as an ADHD coach” prompt. Role-play prompts change the tone of the answers and nothing else. Tone was never my problem.

No “be concise” prompt. Concision is a formatting preference. If output length is your issue, a formatting rule belongs in your settings, not in every message. Developers have a communication skill for this; it is a different project from mine, and I explain the difference in the install guide.

No motivation prompts. A tool that tells you that you can do it is a tool you will stop believing by Thursday.

Where these came from

Every prompt above is the plain-text form of a rule in the file my Claude Code setup reads at the start of every session. The rules have names there: the One Question Rule, Evidence-First Completion, Async Checkpoint Updates, Task Atomicity. If you write code, putting them in a file means you never have to type them, and that is the whole reason my setup holds on bad days. Start with the ADHD developer’s CLAUDE.md guide.

If you do not write code, the prompts are the system. The habits around them, energy-aware planning and the forgiveness rules, are in the ADHD engineer’s productivity system, and the no-code walkthrough is how to use Claude for ADHD.

The whole thing, packaged so you can install it in an afternoon, is the ADHD Productivity System.