Article

How to use AI to start a task you keep avoiding

By Lachie · CyberFreedom · Updated 29 July 2026

Don't ask AI to break the task down. Ask it to write the first three minutes of the work as if it were you, badly, in your voice. The moment that rough draft exists, paste it into the real document and fix it. You're now editing instead of starting, and editing is the easy version of the same task. That's the whole move. It works because it skips the exact step your brain won't do, the bridge from "I intend to start" to "my hand is moving," instead of nagging you across it.

Here's the longer version, including the prompt and the ways it goes wrong.

Why can't I just start, even when I want to?

Because starting a task isn't one action. It's a stack of micro-decisions (where do I begin, what's the first move, what does "done" look like for the first chunk) and every one of those draws on the same depleted executive-function battery. For an interest-based brain, the activation energy to cross from intention to action isn't reliably on tap. It floods in when something's novel or urgent enough, and it's just absent the rest of the time, no matter how much you want it there.

It's 9am. I know what to do. I want to do it. I've had the coffee. And I open the file, read the first line, notice I'm thirsty, close the file, open an easier one. By 10:30 I've done four small adjacent things and not the one thing. By 11 the shame arrives, and shame is the most reliable way to keep me frozen. The same task that was "plenty of time" at 9 is a five-alarm fire at 4. Nothing about the task changed. Only the battery, plus the shame stacked on top.

Task initiation is the gap between knowing what to do and doing the first physical action of it. For some brains that gap is a small step you don't notice taking. For ours it's a wall.

So this isn't a willpower problem. "Just start" names the exact thing your brain won't produce. Telling someone with a task-initiation deficit to just start is like telling someone with a vision deficit to "just see better." The willingness was never the missing input.

Why doesn't "break it into smaller steps" work?

Because breaking a task down is itself an executive-function task. It needs the same planning, sequencing and deciding that's already offline. When someone tells you to break it down, they're asking the broken tool to repair itself. On a good day you can do it. On a frozen day, "break it down" just adds a second thing you can't start to the first thing you can't start.

Same story with the other standard fixes. Willpower works for people whose activation energy responds to being pushed and does nothing when the battery is flat. Streaks and discipline systems work until the first missed day, then become another source of shame. Time-blocking gives you a beautiful calendar full of blocks you'll stare at and not begin. None of these are stupid. They're built for a brain that produces activation energy on request, and that's not the brain that keeps avoiding the task.

What is the actual move, then?

Make something outside your own head do the initiation, and have it hand you back a state where the gap is already crossed. AI is unreasonably good at this, because the thing it does best (generate a concrete first draft from a vague intention) is exactly the thing your director can't do when it's offline.

The core technique: don't ask the AI to break it down. Ask it to write the first three minutes of the work as if it were you.

Here's the prompt I actually use:

I have to [task] but I can't get started. The actual blocker is [X]. Don't tell me to break it down, I know that doesn't work for me. Instead, write me the first three minutes of the work as if you were me. Make it concrete. Make it boring. Make it impossible to misread.

Every line earns its place. "Don't tell me to break it down" stops the model defaulting to the standard advice. "As if you were me" asks for real first-draft work, not instructions about the work: a half-written opening paragraph, the first lines of the function, the actual first sentence of the email. "Concrete, boring, impossible to misread" strips out anything you'd have to interpret, because interpretation is a decision, and decisions are what froze you.

Two things lower the bar even further. First, if you can't even state the task cleanly, don't try. Tell the model "ask me questions until you understand what I'm trying to do" and let it interview you. Answering one question at a time is easy mode. Composing a clean brief from a blank page is the hard executive-function task you're avoiding in the first place. Second, stop typing. I run almost all of this by voice, dictating with a voice-to-text tool, because talking gets the thoughts out with far less friction than staring at an empty field. For an ADHD brain, talking is often the lowest-effort way in.

What you get back is not a plan. It's a running start, real work already in motion that you only have to continue. Continuing is a different cognitive act than starting, and for our wiring it's the easy one.

How do I actually run it, step by step?

1. You type the prompt with your real task and real blocker filled in. Or you dictate it.
2. The AI writes the first three minutes as you. The opening paragraph, the first code block, the literal first sentences. Probably a bit wrong.
3. You don't reply. You take what it wrote, paste it into the real document, and edit it. You are now working. The freeze is broken, not by willpower, but by having skipped the part your brain can't do.

The 90-second rule: aim to be in the work within 90 seconds of running the prompt. If the AI hands back something you have to think hard about before you can use it, the prompt was too vague. Add more constraint and run it again. The output should be so concrete that using it requires no decision. You should be editing, not evaluating.

A real one: I had to write the foreword for my book. Most exposed thing in it, and for a week I opened the file every morning and closed it. I ran the prompt with that as the task and "I freeze at the blank page" as the blocker. What came back wasn't right. Voice slightly off, one detail wrong. Didn't matter. I now had two paragraphs that were 70% me and 100% wrong in ways I could see, and fixing wrong things I can see is something my brain does effortlessly. I rewrote both in five minutes, kept going, had a draft that day. It didn't write my foreword. It got me unstuck enough to write my own.

What if it still doesn't work?

A few common failure modes, with the fix:

• You start improving the prompt instead of the work. The over-engineering trap in disguise. You'll feel productive tuning the perfect first-three-minutes and write nothing. I've surfaced hours later with an elaborate multi-AI machine and zero words on the page. Fix: the prompt is good enough the first time. Take what it gives, even if mediocre, and edit it in the real document. The output's job is to be edited, not to be good.
• The task is too big for "the first three minutes" to land. If the task is "launch the product," there's no coherent first three minutes. Fix: ask the AI "what is the smallest first deliverable inside this?" then run the first-three-minutes prompt on that. You're still not breaking it down yourself.
• You freeze on even opening the AI. Some mornings the freeze is that deep. Fix: lower the bar to the floor. A saved prompt, one keystroke, no typing.
• It works and you don't trust it. You'll suspect you cheated, that real work should feel harder to start. Let it go. You used a prosthetic to do a thing your brain doesn't do. That's no more cheating than glasses are. The work shipped. That's the only score that counts.

Where to take this next

The fastest test is to run it tomorrow morning when you hit the wall. The prompt above lives in the free Prompt Library on the site, ready to copy, so you don't have to retype it from here.

If crossing the gap feels different from being nagged across it, the full thing is the ADHD AI Operating Manual: this workflow plus five more, one per failure mode, every prompt and every honest failure case included. Pay what you want from $9. I wrote it using these exact workflows, which is genuinely the only reason it got finished.