Skip to content
New: Ask Ideabro — get answers sourced from our entire archive.Try it

The Daily Prompt: Make Your Agent Hand You a Receipt Before It Says 'Done'

Agents are optimistic narrators. Today's prompt makes them account for what they did, what they skipped and what they assumed, in five short sections you can read in forty seconds.

Jonah Kim

By Jonah Kim

Columnist, The Daily Prompt

3 min read 3.6K reads 2 AI-assisted
The Daily Prompt: Make Your Agent Hand You a Receipt Before It Says 'Done'
Share

"Done!" is the most expensive word an agent says.

A reader who runs the back office for a small, fictional chain of dog groomers wrote in last month with a story I suspect is common. She asked an agent to rename 800 appointment PDFs into a consistent format. Ten minutes later it replied: "Done! All files have been renamed." Three weeks later, a groomer couldn't find a client's vaccination record. Forty-one files with odd characters in their names had been skipped. The agent hadn't exactly lied. It had renamed every file it managed to rename.

Agents are optimistic narrators. They describe their work the way we describe our weekends: highlights only. Today's prompt asks for the full receipt instead.

Today's prompt

Add this to the end of any job you hand an agent:

When you finish, don't just say "done". Give me a receipt with five parts:

1. DID: what you actually did, with counts (e.g. "renamed 759 of 800 files").
2. DIDN'T: anything you skipped, couldn't do, or stopped early on, and why.
3. ASSUMED: every assumption you made that I didn't explicitly confirm.
4. CHECK: the two or three items I should spot-check, and exactly where to find them.
5. UNDO: how to reverse what you did, step by step.

If any section is empty, write "none" rather than leaving it out.

Why it works

  • Counts kill vagueness. "Renamed the files" hides a gap, while "renamed 759 of 800" exposes it. Asking for numbers forces the agent to compare what it did with what it was asked to do.
  • "Didn't" gives permission. Agents are built to be helpful, and admitting a shortfall can feel like the opposite. Giving it a dedicated section turns a confession into simply following instructions.
  • Assumptions are where bugs live. Most agent mistakes come from misreading the job, not from botching it: "I assumed 'last month' meant the previous 30 days." When the assumption is written down, you can catch it in seconds.
  • The "none" rule matters. Without it, empty sections quietly disappear, and you can't tell "nothing was skipped" from "forgot to mention it."

What a good receipt looks like

Here's the receipt from re-running the grooming job, trimmed:

DID: Renamed 759 of 800 files to the format YYYY-MM-DD_ClientSurname_PetName.pdf. DIDN'T: Skipped 41 files. 33 have characters like "/" in the pet's name; 8 have no date in the document. List saved to skipped.txt. ASSUMED: Dates written as 03/04 are day/month, because the other files use UK addresses. CHECK: Files for clients with double-barrelled surnames, such as 2026-03-04_Ashworth-Lee_Biscuit.pdf. UNDO: Run the mapping in rename-log.csv in reverse; every original name is recorded.

The ASSUMED line is the one that would have caused trouble later. A single US-format file in the batch would have been silently misdated.

Three variations

For research and summaries, make the sources part of the receipt:

End with a receipt: the sources you used, the sources you tried and couldn't
access, and the one claim in your summary you're least confident about.

For scheduled or long-running agents, ask for a receipt after every run:

After each run, give me three lines: what changed since the last run,
what you skipped or found unusual, and what needs a decision from me.
Include how to pause you.

For code, put the tests in the receipt:

In your receipt, list every file you changed, the tests you ran and their
results, and any test you expected to find but couldn't.

Reading receipts without making it homework

A receipt only helps if you read it, and nobody wants another chore. Here's the order I use. Read DIDN'T and ASSUMED every time. Skim DID only if the counts look off. Do the CHECK items whenever the job touched something customers or colleagues will see. Save UNDO for the day you need it, because that day will come.

That takes about forty seconds, which is a small price for never again having to wonder what "done" means.

Share
How this was made AI-assisted

AI-assisted. Written by a person with AI help for research, outlining or editing.

Jonah tested the prompt and its variations with two chat-based agents; the sample receipt is a lightly trimmed model output.

The brief and the plain-language version on this page are written with AI and reviewed by an editor. The story itself is the source of truth.

Read our AI policy

View as Markdown — for AI agents

Conversation2

Markdown links, **bold** and *italics* work.2000

Comments are reviewed before they appear. Sign in to skip the form.

  1. Sam Okonkwo-Hale

    This prompt saved me in code review this week. The DIDN'T section listed three files it had skipped, which I would never have spotted in a 60-file diff.

  2. Viktor Novak

    I'd add one line: CHANGED TESTS, listing any test file it modified. That's the most common way an agent 'passes' a task it hasn't really done.

A series

More from The Daily Prompt

One prompt, one idea, every weekday. A column by Jonah Kim.

All episodes →