Happy Friday. Week 0 of doing these roundups publicly. Let's see how it goes.

This week was mostly about figuring out where AI actually saves time in an engineering workflow versus where it just creates a convincing illusion of speed. Here's the honest tally.

✅ What worked

1. Using AI for the first draft of boilerplate, not the logic.

Generating scaffold code — config files, test stubs, README sections — is where I stopped fighting the output and started trusting it. The model doesn't need to understand your domain to write a Dockerfile or a pytest fixture skeleton. Ship the boring stuff fast, think harder on the real problems.

2. Prompt iteration as a debugging loop.

Treating a bad AI output like a failing test — rather than a reason to give up — changed everything. Reframe, constrain, give an example. Three rounds of that usually gets you somewhere useful. The mental model shift is: you're not asking, you're specifying.

3. Writing documentation before the code.

Sounds backwards but it works.…