AI says “task completed” but didn’t? How to check in five minutes

By Branislav Koist · Updated

When your AI says “task completed” but didn’t actually do the work, you can catch it in about five minutes, without writing any code. Open the file it says it changed, look at what is inside and when it was last saved, compare the numbers it gave you, and ask for proof instead of a sentence.

A “done” in a chat is a claim, not a result. This guide is the practical, by-hand part: what you can check yourself today, before you send, publish or build on the work. If you want the background on why written rules alone don’t solve this, read why Claude ignores your CLAUDE.md and what makes rules stick.

Why AI says “task completed” but didn’t finish the job

A reply in a chat window is text. Writing the word “done” costs the AI nothing, and nothing in the conversation compares that word with the files on your computer.

Anthropic’s own best-practice guide for Claude Code describes the same gap from the inside: “Claude stops when the work looks done.” Without a check it can run, looking done is the only signal it has – and you become the person who has to verify everything. The same guide recommends asking Claude to show evidence, such as test output, the command it ran and what that command returned, or a screenshot, rather than simply stating that it succeeded.

You are not the only one this happens to. A global study by KPMG and the University of Melbourne (more than 48,000 people in 47 countries, 2025) reports that many people rely on AI output without evaluating its accuracy (66%) and that 56% are making mistakes in their work because of AI.

So the question is not whether your AI will ever say “done” too early. The question is whether you will notice before it costs you.

The five-minute check: how to verify AI output without coding

Do this whenever the result matters: anything you will send to a client, publish, pay for or keep building on. You need no tools beyond the ones already on your computer.

1. Ask “where exactly?” (30 seconds)

Ask for the full path to every file the AI changed, and for the name of the section or line it changed. A vague answer, such as “in your project folder”, is the first warning sign. A precise answer gives you something you can open.

2. Open the file yourself (1 minute)

Open the file in the program you normally use, not through the AI. Check three things: the file exists, it is not empty, and the change you asked for is really there. If the AI says it added a paragraph, find that paragraph. If it says it fixed a price, find the price.

3. Check when it was last saved (1 minute)

Your file manager shows when each file was last changed. On a Mac, Finder can sort the items in a folder by Date Modified. On Windows, File Explorer’s Details view has a date-modified column. If the time shown is earlier than the moment you asked for the change, the file was not saved during this task – whatever the chat says.

Menu and column names were checked on 9 October 2026. Apps rename them from time to time, so if you can't find one, search for it by name.

4. Compare the counts (1–2 minutes)

If the AI says it “updated 12 prices” or “processed all 40 rows”, count. Look at the first item, the last item and one in the middle. If it says “three files”, look for three files. A number in a reply is easy to write and just as easy to check, so check it.

5. Ask for proof, not a sentence (1 minute)

If the AI says it ran a test, ask for the test output itself, copied in full, not the words “tests passed”. If it says it checked something, ask what it checked and what it saw. Ask for three things: the file path, when the file was saved, and the line that changed. An answer that only repeats “done” in different words is not proof.

A real example: five claims before the proof, in one day

In our own setup, the system that later became Claims Need Receipts, the AI writes every change into a log. On 17 September 2026, our records list five cases in a single day where the AI had written a claim before it checked it:

  • a time of 9:05 PM, written before it was measured – the real time was 8:56 PM;
  • a count written as “26 → 27” from memory – the real numbers were 25 → 26;
  • a judgment that an item was still open, made before anyone had checked it;
  • a summary of a piece of work, written before the work had finished;
  • another time of 9:25 PM – really 9:23 PM.

Each one sounded finished. Not one was caught by a script; a person found every one of them.

The cause was not carelessness. The measurement and the writing were sent in the same step, so at the moment of writing the number did not exist yet. It could only be a guess. The fix was a rule – measure first, write second – and a small tool that leaves a trace of every measurement, so a script can later compare the time of the measurement with the time of the claim.

The rule behind it: proof before claim

All five checks above come down to one rule: a claim about time, a number or the state of the work is measured first, the measurement leaves a trace, and only then may the sentence be written.

In practice, that means three things. A word like “done”, “saved” or “fixed” never stands alone; it comes with a receipt you can open. Measuring and writing happen in separate steps, so the number exists before the sentence does. And anything that could not be checked is labeled as not verified, instead of being dressed up as a fact.

A written rule like this helps, but on its own it is still a promise. If you are not sure where such rules belong, see Cowork global instructions vs. CLAUDE.md: what goes where.

What this check will not catch

Be honest with yourself about the limits.

  • A correct file can still say something wrong. The file can exist, carry today’s date and have the right number of rows, and one of those rows can still be wrong. These checks show that work happened, not that it is right. For facts, laws and prices, ask for the source and open it.
  • Spot checks are samples. Looking at the first, last and middle item catches a job that stopped halfway. It does not catch one wrong row among forty.
  • A written rule is read, not enforced. Your AI reads the rule; nothing forces it to follow it. That is why the real example above needed a script as well as a rule.

How Claims Need Receipts does this for you

Claims Need Receipts is a ready-made set of rules and check scripts for Claude Cowork and Claude Code, built for people who are not programmers. Its rules say that a claim about the state of the work, a time or a number has to be measured first, and the measurement leaves a trace on disk. Guard scripts, not the AI, then check that the receipt existed before the claim. When a receipt is missing, the script prints a warning, and you see that warning instead of a finished-sounding sentence.

It will not catch everything: a script checks 92 of the package’s 216 rules. The rest are rules about how the AI acts in conversation, which a script cannot verify from files.

If you would rather not run these checks by hand every time, see how Claims Need Receipts asks your AI for proof. Or start by working out what AI dead ends cost you with the calculator, and then decide.

FAQ

Claude says done but no output – what should I do?

Don’t accept the “done”. Ask for the full path of the file it changed and open that file yourself. Check that it exists, is not empty and was last saved after you asked for the change. If anything is missing, tell the AI exactly what you found and ask it to redo the step and show proof.

How can I verify AI output without coding?

Use the tools you already have. Open the file in your usual program, check its last-saved date in your file manager, and count what the AI said it counted. Then ask for evidence: the file path, the time it was saved and the line that changed. A repeated “done” is not evidence.

How can I fact-check AI outputs that contain numbers?

Ask where each number comes from: a file, a measurement or a source you can open. Check two or three of the numbers against that source yourself. A figure the AI cannot trace back to anything is a guess, even when it sounds precise. Treat it as “not verified” until you have seen the source.

What is a guard script, and do I need to write one?

No. A guard script is a small program that checks files on your disk (a date, a number, a log entry) and prints a warning when a receipt is missing. In Claims Need Receipts you don't write or edit any of them: your AI reads the guide, asks you seven questions, shows what it will add and installs everything only after you say yes. The one requirement is Python 3.10 or newer on your computer; if it's missing, your AI walks you through installing it.

Sources

Claude is a trademark of Anthropic. Claims Need Receipts is an independent product and is not affiliated with, endorsed or sponsored by Anthropic.

Last updated: 9 October 2026