Agents as collaborators, not autocomplete

For about a year I treated coding agents like a fancier autocomplete — type a comment, get a function, glance at it, accept. It was fine. It wasn't transformative.

The shift came when I stopped writing prompts at the agent and started writing prompts to it.

What changed

Three small things:

• Brief like a colleague. Why does this matter, what have I tried, what am I worried about. Terse imperatives produce shallow work.
• Hand over context, not instructions. A good prompt proves I understood the problem. The agent fills in the work, not the thinking.
• Trust, then verify the diff. "Done" from the agent means "intended to do" — never "actually did."

A small example

I had a flaky test in a Go service. Old me would have written:

Fix the flaky test in payments_test.go

Now I'd write something like:

Fix the flake in TestRefundIdempotency. Context: it's racy on the
ledger write — two goroutines compete for the same row and the
test asserts a specific ordering. I want the test to assert
the invariant (one row exists with the final state), not the
ordering. Fix the test, not the production code.

The second version doesn't take longer to write. It just front-loads the thinking that I'd otherwise do after reading the agent's first attempt.

The trade

Briefer prompts feel faster, but you pay later — re-reading bad output, re-prompting, undoing. Investing 30 seconds in the prompt saves 5 minutes of cleanup. Compounds across a day.