Two attempts — isn't that a bit harsh?

It is a threshold, not a ban. Nothing stops you retrying a third time when the second clearly got closer. The rule targets the other case, far more common: three attempts that look alike and an hour that disappears.

If I stop, do I lose what was done?

No, provided you restart from a clean state rather than letting attempts pile up. That is exactly what checkpoints are for: returning to the moment before the attempts, without untangling what each one left behind.

And if the problem stays unsolved?

It was already unsolved after two attempts. What you gain by stopping is a clean project and a precise description of the problem — the two things that let you pick it up tomorrow or hand it to someone else.

How many times to retry the AI before you stop

Third retry. It sets off with the same confidence as the first, and you can already tell it will lead nowhere. But stopping now means admitting the last hour is gone. So you retry once more, because maybe this time will be the one that finally works.

Two attempts on the same problem, not three. Beyond that, an AI is no longer searching: it shuffles variants of what it already tried. The signs are clear — it stops rephrasing the diagnosis, it returns to a file it already fixed, it promises exactly what it promised before.

Why we retry one time too many

It is not naivety. It is that each retry is cheap — one sentence, a few seconds — while giving up seems to cost everything already invested.

Except the maths is wrong. The first hour is spent either way. The only honest question is: does the next attempt stand a better chance than the last one? After two failures, the answer is almost always no, and for a simple reason.

What changes between the second and third attempt

On the first two attempts, an AI genuinely explores: it forms a hypothesis, tests it, learns something from it.

By the third, it has exhausted the hypotheses available from the context you gave it. It is no longer searching — it is shuffling variants of what it already tried. From your side it looks like work, with files scrolling past and confident explanations. In reality, nothing new is entering the equation.

Nothing new will enter until you supply information it did not have.

The three signs that never lie

It stops rephrasing the diagnosis. Early on it said “the problem probably comes from X”. Now it writes “let me try something else”. The disappearance of the hypothesis is the most reliable signal there is.

It returns to a file it already fixed. Same spot, third wording. It is going in circles.

It promises exactly what it promised before. Almost word for word. If the announced conclusion is identical while the result never has been, the next attempt ends the same way.

Any one of these is enough to stop. All three together and you have been looping for a while.

What to do instead

Restart from a clean state. Three attempts have left traces in files you were not tracking. Going back to the point before them takes a second and saves you from untangling what each one left behind. Tools keep track of those steps precisely for this.

Change the question, not the wording. Instead of “fix this bug” for the fourth time:

“Do not fix anything. Explain what is happening, in plain language, and tell me what you would need to know for certain.”

That sentence unblocks more situations than any retry, because it leaves repair mode and returns to diagnosis. Often the answer is “I would need to know what you actually see on screen” — information you have and it does not.

Check that it is looking in the right place. A symptom surviving three genuine fixes rarely comes from where everyone is looking. The file being edited may not be the one producing what you see.

And if nothing unblocks, write it down and move on. A problem described precisely — what you expected, what you get, what was tried — takes ten minutes to pick up tomorrow. The same problem abandoned in a mess takes an hour.

Making it automatic

One line in the instruction file your tool re-reads at every start saves you from deciding in the heat of the moment:

Two attempts maximum on the same problem. Beyond that, stop, tell me what you tried and what it produced, and wait.

The number does all the work. “Do not get stuck” means nothing to a machine; “two attempts maximum” can be checked.