Most of the friction people hit with AI coding agents isn't the model — it's how we hand work to it. After pairing with agents daily for months, three small habits have made the biggest difference for me.
1. Give the agent a way to verify itself
An agent that can run the tests, read the output, and try again is worth far more than one that only emits code. Before asking for a change, make sure there's a command it can run to know whether it succeeded. "Make the test pass" beats "write this function" almost every time, because success becomes something the agent can check instead of something you have to.
2. Keep the task smaller than you think it should be
Large, vague asks produce large, vague diffs that are hard to review. Scoping a request to a single file or a single behavior keeps the change reviewable and the feedback loop tight. You can always chain small wins into a big one — but you can't easily un-tangle one sprawling change.
3. Write down the intent, not just the instruction
Telling an agent what to do gets you one result. Telling it why — the constraint you actually care about, the thing that must not break — lets it make the dozens of small decisions you never spelled out. Intent travels further than instructions.
None of this is tied to a particular tool. It's about treating the agent like a fast, literal collaborator that does its best work when the goal is concrete and checkable.
Top comments (0)