How to Use AI to Plan Your Week

Ask a model to plan your week and it will produce something beautiful and useless: eleven hours of focused work in an eight-hour day, no travel time, no lunch, and every task estimated at exactly the length that makes the arithmetic work.

This is not a prompting failure. It is the model doing what it was trained on. The internet's supply of weekly plans is overwhelmingly aspirational — plans as they were written on Sunday, not as they survived Wednesday. Ask for the genre and you get the genre.

Getting something usable means fighting that default deliberately, and the fight is mostly about capacity.

Give it your real capacity, not your working hours

The single change that makes AI planning work: tell it how much time you actually have, after everything already claimed.

Not I work 9 to 5. That's 40 hours, and you do not have 40 hours. You have 40 hours minus standing meetings, minus the commute, minus the hour after lunch when you're useless for hard thinking, minus the two hours a week that go to things nobody schedules. Perhaps 22 hours of genuinely allocable time, of which maybe 10 are good for demanding work.

Give it the real number and add the instruction that matters:

This is my actual available time. Do not exceed it. If the work doesn't fit, tell me what won't fit rather than compressing estimates to make it fit.

Without that second sentence, an over-full week gets silently reconciled by shrinking every estimate. You end up with a plan that fits on paper and fails by Tuesday — the worst outcome, because it looks like you failed rather than the plan.

An honest these three things don't fit; here's what I'd cut is worth more than any schedule. It gives you the decision to make, which was the actual problem.

Brain dump first, structure second

The most reliable planning workflow isn't asking for a plan. It's dumping everything in your head, unsorted, and asking for structure.

Everything: the deadline on Thursday, the thing you keep forgetting, the vague sense you owe someone an email, the appointment you haven't booked. Unordered, unfiltered, badly written. Then:

Group these into themes, flag anything with a hard deadline, identify what's genuinely urgent versus what feels urgent, and tell me what I've listed that isn't actually a task.

That last clause earns its place. Brain dumps are full of things that aren't tasks — worries, decisions waiting on someone else, items that are really three tasks in a trench coat. Separating them is most of the relief, and it's a sorting problem, which is what models are good at.

This works because the hard part of planning is rarely sequencing. It's getting a complete picture out of your head, where things are stored by anxiety rather than by priority.

Estimation is where it fails, and where you help it

A model has no idea how long your work takes. It will guess from the average of all similar-sounding tasks, and the average is wrong for you specifically.

Supply your own estimates, or ask it to estimate and then correct the ones that are wrong — the corrections carry forward within a conversation. Better still, tell it your multiplier. Most people take roughly twice as long as they predict, and if you know your factor is 1.8, say so and let it apply that to everything.

The consequence: a model can help you arrange work but not predict it. Anyone selling automated estimation is selling you the average of strangers.

On the 23-minute statistic

You'll meet this claim in nearly every productivity article: it takes 23 minutes and 15 seconds to recover from an interruption. It's always attributed to a 2005 CHI paper by Mark, González and Harris.

It isn't in that paper. We ran the PDF looking for it — the string doesn't appear. The paper reports that 57% of working spheres are interrupted, and that interrupted work usually resumes the same day but with more than two intervening activities first. The 23-minute figure comes from a 2006 Gallup Business Journal interview with Gloria Mark. The finding is real; the citation almost everyone uses is not.

The practical point survives regardless: context switching is expensive, and batching similar work is the highest-value structural change you can make to a week. Ask a model to group your tasks by the kind of attention they need — deep focus, shallow admin, conversations — rather than by project. Three hours of admin in one block costs less than the same three hours scattered across six interruptions.

What models are genuinely good at here

Turning a vague goal into a sequence. Launch the new pricing page becomes eight concrete tasks with dependencies. The vagueness is the blocker; decomposition is exactly the model's strength.

Prioritising against stated criteria. Give it a framework — urgency against importance, or reversibility, or whatever you actually use — and a list, and it will apply the framework consistently. Consistency is the value; you'd apply it unevenly at 4pm on Friday.

The weekly retro. What got done, what didn't, what pattern connects the things that didn't. Reflection questions are a genuinely good use because the model has no stake in the answer and won't let you off.

Naming the real blocker. Describe a task you've avoided three weeks running and ask what's actually in the way. Often the answer is that it's under-specified, or it depends on a decision you haven't made. Both are fixable; procrastination isn't.

The productivity and planning prompts here are all built around that realism constraint — respect actual capacity, never over-schedule, and be honest when something doesn't fit rather than producing a fantasy that collapses on contact.

What to be careful about

A weekly plan is not a commitment device, and the neat output can feel like progress in itself. If planning is where your week's energy goes, more sophisticated planning is not the fix.

And the model doesn't know what matters. It will optimise whatever you give it, which means a well-arranged week of work on the wrong things looks identical to a well-arranged week on the right ones. That judgement stays yours, and no amount of structure substitutes for it.

The short version

1. Give it your real available hours, not your nominal ones
2. Forbid compression — make it tell you what doesn't fit
3. Brain-dump first, ask for structure second
4. Supply your own time estimates, or your personal multiplier
5. Batch by attention type, not by project
6. Use it hardest on decomposition and retros, where it's strongest