1 min lesson
Minutes 15-70: build the slice, narrate the AI
Use "Minutes 15-70: build the slice, narrate the AI" to tell the cases apart, then choose a response for each one.
Step 1 of 2
Minutes 15-70: build the slice, narrate the AIThe AI-authenticity test is live here
Boilerplate scaffolds: table DDL, repository functions, test harness.
Well-trodden syntax you can verify at a glance.
Say: 'accepting this, it's standard and I can read it.'
Money math the model guessed: rounding, currency as float, recognition timing.
Anything that mutates a closed period or skips the idempotency key.
Say: 'rejecting - it used a float for currency; I'm switching to integer minor units.'
When the AI's first cut has a bug - it will - do not hide it. Catch it, say what is wrong and fix it. “The model wrote the dedup as DISTINCT over all columns, but two genuine events can share every field except event_id, so I'm switching to ON CONFLICT (event_id).” Visibly debugging AI output is a stronger signal than code that happened to come out clean, because it proves the judgment they cannot interview for any other way.