FIX
Twenty prompts in, and the same bug is still there.
Past three failed attempts, more prompting tends to add damage rather than remove the bug. Humanized Code breaks the loop: a senior engineer reproduces the failure once, repairs the cause instead of the symptom, and writes it up with a fixed repair quote inside 48 hours, for $50.
01
The shape of the loop
It has a recognizable rhythm. The model apologizes, proposes a confident fix, and the error changes slightly. You paste the new error. It apologizes again and reverses part of the previous fix. Four rounds later the first bug is back and two new ones have joined it, and the file no longer resembles anything you would have written.
The worst part is not the wasted hours. It is that after fifteen edits you can no longer tell which version was closest to correct, so even starting over has become expensive.
02
Why more prompting stops helping
The model is reasoning about your code without being able to run it. It cannot see the request that failed, the row that was missing, or the environment variable that was blank. It is inferring the cause from the shape of the code, which works beautifully for a class of bugs and not at all for the ones that survive three attempts.
It also has no memory of the two fixes before last, so it will happily propose one of them again in different words. And it is built to produce an answer — "I would need to see the actual request payload" is a more useful reply than a fifth guess, but it is not the reply the format encourages.
03
A stopping rule you can use tonight
If three consecutive attempts have not changed the error message, stop. Save the current state somewhere you can get back to, write down the last version you know worked, and change approach — either gather the missing evidence (the real log, the real request) or get a person involved. This rule is free and it is worth more than most of what we sell.
$50 FLAT — FINDINGS + A FIXED QUOTE IN 48 HOURS
$50
Flat initial review
48h
Written findings + fixed quote
START THE REVIEW
- How many attempts is too many?
- Three without the error message changing. Movement means you are converging; the same failure in new wording means the model is guessing and will keep guessing.
- Should I delete it and start over?
- Rarely worth it. The bug you cannot find is usually one wrong assumption, not a bad codebase, and a rebuild reproduces the same assumption faster.
- Will you just run it through another AI?
- No. A senior engineer reads your code and reproduces the failure, and that is the entire product. You have already tried the other thing, at length, which is why you are here.
- What do you need from me to break the loop?
- The repo, the exact steps that trigger it, and the error text. If you have the chat transcript, send it — knowing what has already been tried saves us an hour.