Observations: Coding with AI

Short observations from coding with AI (Cursor, Claude, Copilot, etc.) in 2024–2026.

Context is the main lever. The model is only as good as what it can see. Open the right file, paste the error, or point to the doc. "In this file, add validation for the email field" beats "add validation."

Small steps beat big prompts. "Refactor this function to use async/await" and then "now add error handling" works better than "refactor this whole module and add error handling and tests" in one go.

The model doesn’t run the code. You do. So run it after every non-trivial change. When it breaks, paste the error and the relevant snippet. The next response is usually a fix.

Naming and structure help. Clear file and symbol names and a predictable project layout make it easier for the model (and you) to navigate. A short README or AGENTS.md pays off.

It’s good at boilerplate and patterns. Repetitive code, standard CRUD, and "do it like the rest of the codebase" are where it shines. Weird business logic and one-off design decisions often need your judgment.

You’re still the architect. The model suggests; you decide. Scope, data model, and UX are your job. The model is a fast draftsman.

Verification is non-negotiable. Review diffs. Run tests. Glance at security and dependencies. Don’t ship on trust alone.

Coding with AI is collaboration, not replacement. The skill is directing the collaboration well. See 9 Things for Agent-Driven Development for a structured list.