Before the AI starts, write a check that fails today and passes only when the job is truly done. Then run that check yourself, or with a small script. Never ask the AI whether it has finished, because “yes” is the easy answer.
That one habit separates a job that is done from a job that has only been reported as done. Here is how it works, and what to put around any job you leave running while you do something else.
Try it in two minutes
Say you ask an AI to tidy a product description so that it contains no em-dashes (the long dash). Before it starts, open the text in your editor and search for that character. The search gives you a number. Today that number is not zero, and that is the point: you now have a check that fails.
The job is done when the number is zero. Let the AI work. When it says “finished”, do not reply “are you sure?”. Run the search again. If it still finds three, the AI was wrong, and you caught it in seconds. If it finds none, the job is done, and you know because you counted, not because you were told.
A good check has three traits:
- It gives a plain number or a yes or no.
- It fails before the work and passes after.
- You can run it without the AI.
If you cannot think of a check, the job is not described clearly enough yet. Fix that first. It is cheaper than finding out later.
Why the AI’s own report is not enough
A report says what the work set out to do. It does not prove what now exists. The two can differ, and the difference can be total.
Here is a real case from our own work. An automated 3D build finished with a log that said it had succeeded. The first picture rendered from that build had the object missing entirely. The log was cheerful and the result was empty.
The lesson is to look at the result, not the log. Render it and look at it. The same goes for any job: open the file, view the picture, count the stones.
Three limits for any job you leave running
If a job runs while you are away, give it three things before it starts.
- A stop. What ends it. Without one, the job drifts on, or quits halfway and calls that done.
- A cap. The most tries, minutes or pounds it may use. Without one, a job that cannot succeed keeps trying and keeps spending.
- A proof. A check the AI cannot mark for itself. Without one, you are back to taking its word.
Take a batch of product photo edits. The stop is “every picture saved”. The cap is “three tries per picture”. The proof is “the stone count in each picture matches the original”. The AI does not tick that last box. You, or a script, count.
Keep a person in charge of the last step
Anything that sends, spends, deletes or publishes should wait for your approval, however good the check looks. A passing check tells you the job is done. It does not tell you the job should go out.