Give a language model your stalled goal and it will produce a plan. A good one, usually — sensible steps, reasonable order, sized about right.
You'll read it, agree with it, and quite possibly do none of it.
That's not a failure of the model. It's a mismatch between what these tools are good at and what's actually wrong with most stuck goals.
What it can genuinely do
Three things, and they're not trivial.
It can break ambiguity, which is the most common cause of stalling and the cheapest to fix. If a goal is a vague noun — sort out the pension, work on the website — a model will convert it into a list of physical actions in about four seconds. That's real work and it's work people reliably fail to do for themselves, because deciding what starting means is the exact task that gets deferred.
It can size things down. Ask for the smallest possible next step and you'll get something smaller than you'd have chosen, because you're not the one who'd have to do it and it isn't protecting anything.
And it can read back what you've written more honestly than you can. Feed it eleven of your own log entries and ask what's missing, and it will notice that none of them touch the difficult thing — which is visible in the record and invisible to you, because you were there for each entry individually.
What it can't do
It can't want the goal, and it can't make you.
That sounds obvious and it's the whole thing. A frightened person doesn't need a better plan; they have a perfectly good plan and can't approach it. Handing them a more detailed one adds another document to a goal that already has too many, and produces a small feeling of progress that displaces the actual move.
It also can't tell you which of the five mechanisms you're in, because it has no access to the information that would distinguish them. Whether dropping a goal would feel like relief or loss is not in your notes. Whether the sessions are absorbing once you start or grim throughout is not in your notes. It can guess from what you've written, and its guess will be shaped by whatever you happened to write, which is exactly the thing distorted by whatever's going on.
And it can't supply consequence. Most of what makes people act is that someone is waiting, or something happens on Friday. An assistant is never waiting.
There's a specific way this goes wrong. Asking a model about a stalled goal is itself a comfortable activity — it's engagement with the goal that carries no risk, produces output, and feels like progress. For a goal stalled on fear, that makes it another form of the sideways work: an evening spent thinking about the problem rather than an evening spent inside it. The tell is a goal with a lot of planning attached and no sessions.
The prompt that returns something usable
If you're going to use one, ask for less than you're inclined to.
Give it the goal, the last few notes, and the blocker if there is one. Ask for three moves of thirty minutes or less, and ask it to quote a line from your own notes so you can check it read them. That's it.
What you're avoiding is the plan-shaped answer — the twelve-step programme with phases, which arrives looking authoritative and functions as one more thing to have and not do. Three small moves is a menu you can act on tonight. A roadmap is a document you'll file.
The quoted line matters more than it sounds. It's a cheap verification that the output is about your situation rather than about stalled goals in general, and if it can't find one, the advice is generic and worth exactly what generic advice is worth.
Where this is unfair to the tools
The counter-argument deserves stating, because there's a fashionable dismissiveness here that's also wrong.
For ambiguity, which is the single most common cause of stalling, these things are genuinely excellent, and dismissing them because they can't fix fear is like dismissing a spanner for not being a hammer. A large fraction of stuck goals are stuck on nothing more mysterious than an undefined next action, and that is a solved problem now.
There's also a real effect from articulating a goal to something that responds. Writing out what's stuck, in enough detail for a model to work with, does some of the thinking on its own — and the reply, even a mediocre one, breaks the loop of turning it over privately. That's not nothing, and people who've experienced it aren't imagining it.
The honest position: useful for the mechanical failures, useless for the emotional ones, and the mechanical ones are more common than the writing on this subject usually admits.
Twenty-two days and three moves
Ship the beta to ten testers, 22 days quiet. The notes say: refactored the db layer, redesigned settings, read about a different framework, decided it should handle recurring appointments first.
Asked for three thirty-minute moves, the model comes back with: send it to one person rather than ten; drop recurring appointments from scope and write that down; and book twenty minutes with the receptionist on Thursday. It quotes the line about recurring appointments and observes that it was invented after the tool already worked.
That's genuinely useful, and none of it is clever. It's the obvious reading of four notes by something with no stake in the outcome — which is precisely what he couldn't do, because he'd written each entry believing it at the time.
What it doesn't do is make Thursday happen. He still has to sit next to someone while she uses the thing he built, which is the part that's been stopping it since August, and no amount of restating the step makes that easier.
He does it, in the end, because the note is on the goal where he can see it and the counter says 22 days. Not because the plan was good.
What the tool is for in goaldrift
One button, on one goal, on request. It reads that goal's notes, steps and blocker and returns three moves of thirty minutes or less, with a line quoted from your own record. Then it's finished.
It doesn't check in, remember your mood, or want to talk. There's a genre of app now where you end up performing for the assistant, filing your day, being told you're doing brilliantly — and that's another thing to keep up with, in an app specifically for people already failing to keep up with things.
It runs on your own key, which stays in your browser. So it costs nothing to run, and the number of your notes anyone else holds is zero.
The counter tells you which goal has gone quiet. The notes tell you what was happening when it did. The model can suggest what to do about it. What none of them can do is the thing on Thursday.
Which is roughly the right division of labour, and worth being clear-eyed about. A tool can remove the obstacles that are made of not knowing and not noticing. Everything else was always going to be yours.
Related: The thirty-minute unstick · What we left out of goaldrift, and why · The anatomy of a stuck goal
Share: