Why is the AI burning my Base44 credits fixing its own mistakes?

Sixty seconds, no AI. Most people are fixed by one of these and never need the prompt below.

1. 1

   Count your last ten messages to the builder. How many were new work, and how many were "that didn't work, try again"? If the second number is higher, the problem is the loop, not the price.

2. 2

   Stop after two failed attempts at the same thing. A third attempt on top of a broken state is the most expensive message you will send, because the AI is now reasoning about code it wrote badly. Revert to the last good checkpoint and re-describe the goal instead.

3. 3

   Check whether you are on the Free plan. Base44's docs: on Free, "both Resolve with AI and automatic error fix messages consume message credits". On paid plans those are free. If you are paying for your own error fixes, that alone can be half your burn.

4. 4

   Use Discuss or plan mode before Build. Planning turns are much cheaper than building turns. Agreeing the approach in discussion and then asking for one concrete change costs a fraction of building, discovering it is wrong, and rebuilding.

5. 5

   Look at what you asked for. "Make the dashboard better" is a request for the AI to guess, then for you to pay for the guess. Name the file, the component and the single behaviour you want changed.

6. 6

   Check status.base44.com if things suddenly start failing. Retrying through a platform incident is you paying for their bad afternoon.

7. 7

   Do not expect a refund. The docs are unambiguous: "Credits are non-refundable for tool behavior and AI mistakes." Refund requests are the most common thread in the Discord and the answer does not change. Prevention is the only lever you control.

Credits are consumed by messages to the builder AI, not by your app running. So your bill is a function of how many times you ask, not how much traffic you get. That single fact reframes the whole problem: the cost driver is your prompting, not your pricing tier.

Then three platform facts make the loop expensive:

You cannot see the cost before you spend it. From the docs: "There is no way to preview an integration credit cost before running an action." A builder in the Discord this week described a project that "churned almost 1k credits in 4 days" after being told only that it would use "an ok amount".

Mistakes are not refundable. "Credits are non-refundable for tool behavior and AI mistakes." If the AI breaks your app and then charges you to fix what it broke, that is the documented behaviour, not a billing error.

Unused credits expire. They do not roll over, and the cycle resets at the UTC time you subscribed — not on your local calendar date. So there is no banking a quiet month.

The compounding problem is the fix loop. The AI makes a change that half works. You say "that's not right". It now reasons about its own broken code, and produces something further from what you wanted. You say "still not right". Three messages in, you have paid three times and are further back than when you started. Reverting to a checkpoint and re-describing the goal costs one message and lands closer.

A Base44 Partner put it plainly in the bug channel this week, replying to someone who had lost half a $50 top-up to error-fixing: the answer is "effective prompt engineering, custom instructions and skills so that the AI does a better job of writing code that works." That is correct, and it is also the part nobody writes down. So here it is.

Paste into your Base44 builder chat. The first one only reads — it changes nothing, so it is safe to run on a live app.

These are standing rules for this project. Follow them on every message from
now on, without being reminded.

SCOPE
Do exactly what I asked and nothing else. Do not refactor, tidy, rename,
reorganise, upgrade a dependency, or "improve" anything I did not mention. If you
believe something adjacent needs changing, say so in one line at the end and
wait. Unrequested changes are how a working app stops working.

BEFORE YOU WRITE ANYTHING
If my request is ambiguous, or could reasonably be built more than one way, ask
me one clarifying question instead of guessing. One question costs a fraction of
a wrong build.
If the change touches more than three files, or changes a data model, tell me the
plan first in under ten lines and wait for my go.

WHEN SOMETHING FAILS
If a change does not work, do not immediately try a different approach on top of
…

Stop. Do not make any further changes yet.

We have been going round in circles on this and I am paying for every attempt.
Before you touch anything else, answer these five questions and nothing else:

1. In one sentence, what was I originally trying to achieve? Not what we have
   been debugging — the original goal.
2. What is the current actual state of the code relative to that goal? Be honest
   about anything that is now worse than when we started.
3. What have we tried, and why did each attempt fail? One line each.
4. What is the single most likely root cause, and what evidence in the code
   supports that rather than the alternatives?
5. Do you recommend we revert to a clean checkpoint and start this change again
   from scratch, or continue from here? Give me a straight recommendation with
   a reason, not both options.

Rules for this answer:
Do not write or change any code in this message.
…

Audit this project for the things that make me spend credits unnecessarily.

READ ONLY. Change nothing. This is a review, not a fix.

Look at the project as it stands and tell me:

1. FRAGILITY — what is most likely to break next
   - Files that are very large, or components doing too many jobs at once. A big
     file means every change to it is a bigger, riskier, more expensive message.
   - Anything with no error handling, so a failure surfaces as a blank screen
     rather than a message — blank screens are expensive to diagnose by chat.
   - Duplicated logic that has to be changed in more than one place, because I
     will forget one and pay to fix it later.
   - Anything that looks half-finished or abandoned.

2. AMBIGUITY — what makes you guess
   - Where in this project would a reasonable instruction from me still be
     ambiguous to you? Name the areas where you would most likely build the
…

• Base44 — credits"Credits are non-refundable for tool behavior and AI mistakes." Also: no cost preview before an action, and no rollover.
• Base44 — billing and plans"All Base44 services are non-refundable per our Terms of service."
• Base44 — changelogCredit rules move. In-app agent replies moved to usage-based charging on 30 Aug 2026, so older per-message figures are wrong.
• Base44 status pageIf the platform is having a bad day, your retries are paying for it.