Writing1 min read

Engineering Notes

AI in the Workflow

Generated code that passed review and failed on null.

The generated code passed review. It failed on a null edge case.

I used AI to draft a parser for optional metadata. Tests covered the happy path. Production rows had null in metadata.items[0].nested.

AI for the boring parts

I use AI for scaffolding, summaries, and repetitive transforms. Speed is real. So is plausible wrong output.

Review trusted the diff

Review treated generated code like trusted input. We skipped the odd paths because the diff looked familiar.

You still own proof

AI assists. I still own verification. Read edge cases, run tests, confirm assumptions.

Better context in the prompt helps. It does not replace proof.

Review checklist

Human-in-the-loop review
Three checks before merge
1. Null, empty, and malformed inputs handled?
2. Tests cover failure paths, not only success?
3. Dependencies introduced are justified and pinned?

Verify like untrusted input

Verify AI output like any untrusted input. Speed without proof is debt with a confident face.