Agentic Workflows That Survive Contact

The useful version of AI-assisted programming is not magic. It is a tight loop: state the goal, inspect the system, make the smallest durable change, and verify the behavior before momentum turns into drift.

I keep a few defaults close. The agent reads before editing. File changes stay narrow until the shape of the codebase is clear. Tests are not a ceremony at the end; they are the next message in the conversation.

Start With The System

Good tools make it easy to ask for a leap. Good engineering asks what the landing zone looks like. Before a refactor, I want to know where the old behavior is encoded, what names the code already uses, and which tests will complain if I get clever in the wrong place.

The best prompt is often not bigger. It is more situated.

Prefer Reversible Moves

Small commits, local helpers, and boring data shapes keep the work recoverable. When a coding agent is moving quickly, reversibility matters more than flourish. A tiny change with a clean verification path beats a broad rewrite that only looks confident.

Make Verification Visible

The moment a test fails, the transcript becomes a map. The failure names the next useful question: is the expectation wrong, the implementation wrong, or the assumption wrong? That is where the real work starts.