The thing that makes it 😱 is:

1. write 1 bad prompt feed it to `planner` that reads the entire project into context and spits out 3 pages of plan
2. automagically send the plan to `coder` that reads the entire project into context and spits out 20k bad lines of code
3. automagically send the PR to `tester` that reads the entire project into context and tells `coder` he's a n00b and sends 50 red herrings as an eff-you
4. `coder` fixes all the 50 red herrings and then send it back to `tester`
5. `tester` pulls the feature branch and needs to load the entire thing into context again. is still pissy too because it doesn't work, back to `coder`: this shit dont work change `<a>-<z>`
6. `coder` runs out of context so needs to compact, gets all the new tokens, fixes the sole bug by accident

So that is 6 times your bill right there.

But then we use `ralphwiggum(n=50)` so it's actually 300x your bill, because every step gets retried 50x

I was wondering how these AI bros were getting big boy bills, but I guess it's mostly by prompting "boil the ocean, I want to take a bath."

![](https://m.stacker.news/130319)

This is just one billing tick, all from this run afaik.

k00b

It updated it's sat filters and E2E tested the search changes and found a bug that I missed!