Ask a language model what to do about a stalled goal and you'll get a plan. Usually a decent one — phases, sequencing, sensible steps.
You'll read it, agree with it, and quite possibly do none of it. That's not the model failing; it's answering the question you asked, and the question was wrong.
Why plan-shaped answers don't get used
Three reasons.
They're too big to start. A twelve-step programme has no obvious entry point on a Tuesday evening. Each step needs sizing, sequencing and a decision about where to begin, and those decisions are exactly what you were avoiding.
They produce the feeling of progress. Reading a good plan delivers some of what doing the work would have delivered. For a goal you're avoiding, that's actively harmful — it discharges the discomfort that might otherwise have produced a session.
They're one more document. A stalled goal frequently has too many already: an outline, a plan, a restructured version of the plan. Adding an authoritative-looking one makes the goal heavier rather than lighter.
Ask for less
The prompt that returns something usable is smaller than the one people write.
Give it the goal, the last few things you wrote about it, and the blocker if there is one. Then ask for three moves of thirty minutes or less, and ask it to quote a line from your notes.
That's it. No strategy, no phases, no timeline.
The constraint is doing the work. Thirty minutes forces the answer down to something startable, and three of them gives you a menu rather than a sequence — you pick one, tonight, without deciding on an approach.
Three is deliberate too. One is a demand and can be refused outright. Five is a list, and lists need prioritising. Three is small enough to read in ten seconds and large enough that at least one will be tolerable on the evening you're actually having.
Why the quote matters
Ask for a line from your own notes, verbatim, and you get a two-second test of whether the response is about your situation.
Without it, models produce competent general advice about stalled goals, which reads as insight and applies to everybody. With it, you can see immediately whether it engaged with the specifics — and if it can't find a line to quote, the advice is generic and worth what generic advice is worth.
It's a small thing that changes how much you trust the output, in both directions. Sometimes the quoted line is one you'd forgotten writing, and seeing it pulled out is itself the useful part.
What the moves look like
Small, specific, and at least one of them uncomfortable.
For a side project stalled at 22 days with notes about refactoring and redesigning: send it to one person rather than ten; drop recurring appointments from scope and write that down; book twenty minutes with the receptionist on Thursday.
Each is doable in half an hour. None requires deciding on an approach. And the third one is the one that's actually been avoided since August, which a model can often see precisely because it has no stake in it.
If all three moves are comfortable, the answer is probably wrong. A stalled goal usually has one uncomfortable next step, and a response that doesn't include it has been too polite.
Where this doesn't help
The honest limits.
For ambiguity — not knowing what the next action is — this works extremely well, and ambiguity is the most common cause of stalling. That's a genuine solved problem now.
For fear, it does very little. A frightened person doesn't need a better next step; they have one and can't approach it. The three moves will be correct and unperformed, and the asking will have functioned as one more evening of engagement that carried no risk.
For a wrong goal, it's worse than useless. A model will help you plan something you should be deleting, enthusiastically, for as long as you keep asking.
And it can't supply consequence. Most of what makes people act is that someone is waiting, or something happens on Friday. Nothing you ask is ever waiting.
Doing it without an app
Nothing here requires a particular tool. The prompt is the whole technique, and it works in any chat interface.
Paste the goal, three or four of your recent notes, and the blocker. Then: Give me three moves of thirty minutes or less that would restart this. Quote one line from my notes so I can tell you read them. No plan, no phases.
The main thing an app adds is that the notes are already there. Assembling the context by hand is the friction that stops people doing this, and it's most of what's automated when it's built in.
What it looks like in goaldrift
One button, on one goal, when you press it. It reads that goal's notes, steps and blocker and returns three moves of thirty minutes or less, plus a quoted line. Then it's finished — no conversation, no memory, nothing accumulating.
It runs on your own key, held in your browser, so the text goes from your machine to the provider you chose and the number of your notes held elsewhere is zero.
What it won't do is the thing on Thursday. That was always going to be yours.
Which is roughly the right division. A tool can remove the obstacles made of not knowing what to do next. Everything else — the wanting, the discomfort, the twenty minutes with the receptionist — was never something software was going to take off you.
Related: AI can't want it for you · The thirty-minute unstick · What we send to the model, and what we don't
Share: