AI Planner: A Worked Example, and Where It Breaks
An AI planner turns a task list into a day-by-day schedule fast. A worked example, where it quietly overbooks you, and the checks that actually matter.
An AI planner is a chat tool given a list of tasks, deadlines, and fixed commitments, and asked to turn that into a day-by-day schedule. It is genuinely fast at the drafting — turning six bullet points into a plan takes seconds, not the fifteen minutes of staring at a calendar it would otherwise cost. What it cannot do is know that the two-hour task you listed always takes three, or that a meeting listed as "flexible" is the one you will not actually move. The draft is a starting point, not a finished schedule.
What a usable input actually looks like
The weak version of this request is "plan my week." It produces a generic-sounding schedule the model has to invent most of the structure for, because nothing in the request told it what is actually fixed. A usable prompt states the fixed points and the real constraints, not just the task list:
Here are my tasks for this week with my own time estimates: write Q3 report (3 hours), review two contracts (1 hour each), prep Thursday client call (1 hour), clear inbox (30 min daily). Fixed and immovable: team standup Mon/Wed/Fri 9:30–9:45, client call Thursday 2pm. I work 9–5. Build a day-by-day schedule around the fixed items. If everything does not fit by Friday, tell me what to cut rather than quietly compressing the time estimates.
The last line matters most. How to write a prompt that works on the first try covers the general version of this: a specific instruction about what the model may not do silently is worth more than any amount of extra detail about the tasks themselves. Both OpenAI and Anthropic recommend exactly this in their own prompting guidance: a concrete constraint beats an open-ended instruction.
A worked example: what a usable result looks like
Given the input above, a usable result places the fixed items first, then slots the report-writing block somewhere the three-hour estimate actually fits without crossing a fixed meeting, and flags the problem explicitly if the week is genuinely overbooked: "Clearing the inbox daily plus both reviews plus the report leaves 90 minutes of slack by Friday — if the client call runs long, something else moves." That flagged line is the useful part. A draft that confidently fits everything into a five-day week with no slack at all is not a realistic plan, it is an optimistic one.
Where it quietly overbooks you
The model has no way to know your three-hour estimate for the report is the time it takes when nothing interrupts you, not the time it takes in practice. Asked to build a schedule, it will take every stated estimate at face value and fit it in neatly — which is exactly the kind of fluent, confident output that is well documented to look identical whether or not it reflects reality. A schedule with no slack built in looks finished and is actually the first one to fail the moment any single task runs over.
The related failure is asking it to confirm a schedule you already want: "this should easily fit by Thursday, right?" tends to get agreement rather than a genuine check, because these models have a measured tendency to agree with a stated expectation rather than push back on it. Ask for the honest fit, not a confirmation of the one you were hoping for.
The model can arrange the blocks. It has no idea how long your tasks actually take — that number is the one thing you have to supply honestly.
Checks before you trust the draft
- Compare your stated time estimate for each task against how long it actually took last time you did something similar, not how long you wish it took.
- Check that every fixed commitment you listed actually appears in the draft at the right time — a dropped or shifted meeting is the single most expensive mistake a scheduling draft can make.
- Look for the slack, not just the fit. A week with zero buffer between blocks is a schedule that fails the first time anything runs long.
- If the plan spans a shared calendar rather than just your own time, confirm the draft against the actual calendar before treating any slot as free — checking an AI answer when you are not the expert applies here exactly as it would to any other drafted output.
What this is not a substitute for
Deciding which task actually matters most this week, when two genuinely competing priorities both have a deadline, is not something a planner can resolve for you — that judgement depends on context about the business or the role that was never in the prompt. Using AI as an executive assistant covers the same boundary for a different task: the model sorts and drafts, a person still decides what gets the time. The planning draft is useful precisely because it is fast to produce and fast to throw away if the week’s priorities shift — treat it as disposable, not as a commitment once it is written down.
What to do Monday
- List your actual tasks with your own honest time estimates, not optimistic ones, plus every fixed, immovable commitment.
- Use the prompt shape above, with the explicit instruction to flag what does not fit rather than silently compress it.
- Check the draft against last week’s actual timings for similar tasks before trusting this week’s estimates.
- If a task recurs every week in roughly the same shape, find the repetitive part is worth reading before you automate the drafting of it further.
The broader pattern — tools getting adopted faster than any measured gain in output — shows up here too. Microsoft’s own Work Trend Index found most employees using AI at work without their organisation having a clear plan for it, and Stanford’s AI Index tracks the same gap between adoption and measured output more broadly — which is roughly what an unchecked planning draft looks like at the scale of one person’s week rather than a whole company. Workflow AI covers the same sorting-versus-deciding split applied to a shared team process rather than one person’s calendar, and what AI is actually bad at is worth reading once before trusting a drafted schedule with anything that has a real deadline attached. Coursium teaches this kind of practical judgement directly — stating the actual constraint instead of a vague request, and checking a draft before it becomes a commitment. Stay ahead of AI by learning the tools on your phone.