How to ask an AI where your time went

A language model is genuinely good at reading a fortnight of time entries and telling you something you had not noticed. It is also perfectly willing to do that with a record that is wrong, and it will sound equally confident either way.

So this page is in two halves: how to get a useful answer, and how not to get a confident one about a fiction.

The thing nobody tells you first

An AI cannot tell how good your data is. It will take a timesheet you reconstructed on Friday afternoon and analyse it exactly as carefully as one recorded to the minute, and the language it uses will be just as certain.

That is not a flaw in the model. It is missing information, and you are the only one who has it. So tell it. Before anything else in your prompt, say how the record was made:

• How many days the record actually covers, out of how many you worked
• Roughly what proportion was written down as it happened, and what was reconstructed
• Where you know there are gaps
• Whether breaks are recorded, or whether they are inside the working entries

That last one matters more than it sounds. A day that shows eight hours with no gaps almost always means lunch got logged as work, and an eight-hour day is a different conversation from a seven-hour one with a proper break in it.

What to paste in

Export your entries as CSV and paste them. Most models handle a fortnight comfortably. What each row wants: the date, the start and end times, the project, the task, and a note if you wrote one. Start and end times matter more than durations, because when the work happened is half of what you are trying to learn.

A prompt that works

The order is doing the work. Asking for the limits first changes what comes after it, because the model has already committed to what it cannot claim. Ask for the analysis first and the caveats arrive at the bottom, after the confident part you have already read.

The questions worth asking

• “When did the real work actually happen?” Most people are wrong about their own best hours, and the record knows.
• “How fragmented was this?” Ask for the longest unbroken stretch and the median session length. If your longest is under ninety minutes, that is the finding.
• “What did I plan, and what got the time?” Only works if you wrote the plan down first — but it is the most useful question on the list.
• “What is suspiciously absent?” Good models are strong at this. The project you swore you were spending your life on, with four hours against it.

And the ones to avoid: anything beginning “am I productive”, which invites a grade, and “how can I be more efficient”, which invites a listicle it could have written without reading your data at all.

What has changed in 2026, and what has not

Connecting an assistant to a time tracker used to mean copying and pasting. It does not any more: during 2026, Daily (June), Timing, Tyme (September), Harvest and ManicTime all shipped MCP connections, which let Claude or ChatGPT query your tracked time directly. Toggl, Clockify and RescueTime have community-built equivalents. Checked 20 September 2026.

That is genuinely convenient and it does not solve the problem this page opens with. A live connection hands the assistant the data; it does not hand over any account of how good the data is, because the tracker does not know either. A timer cannot tell you about the afternoon you forgot it, and a screen-reader cannot tell you which of those logged hours was actually lunch.

Which is why the paragraph you write at the top of the prompt is still the most important part, whether you paste a CSV or plug in a connector.

What to do with the answer

Disagree with some of it, out loud, in the thread. Models fold immediately when challenged on a specific, which is a weakness generally and a useful property here: push back on a conclusion and you will quickly find out whether it came from your data or from the general shape of advice about time.

Then take exactly one of the three changes and run it for a week. Three at once tells you nothing about any of them.