openai_build_prompt() generating too long a prompt

[simonw]

Something seems to have gone wrong here: https://datasette.simonwillison.net/simonwillisonblog/answer_question?question=Use+cases+for+newsrooms+using+Datasette+Cloud&_hide_sql=1

Is my token counting not right? Is there a logic problem with how it decides how much of an extract of each one to sue?

Ideally need to replicate this in a test.