· 4 min
AI speed with engineering taste
AI coding tools can shorten the time between an idea and a first draft. They do not remove the engineer’s responsibility for deciding whether the change belongs in the system. Engineering taste is the ability to choose a small design that fits the codebase, reject output that adds hidden obligations, and explain the behavior well enough for someone else to maintain it.
Use a short control loop
- Define the task before prompting. State the user-visible behavior, constraints, relevant interfaces and what must remain true.
- Ask for a bounded change. Provide the nearby code and request one coherent step, not a broad rewrite with guessed requirements.
- Review the diff as an author. Check data flow, error handling, permissions, logging, dependencies, lifecycle behavior and consistency with local patterns.
- Run the right checks. Use deterministic tests and static checks for ordinary code; separately evaluate any model-dependent behavior using representative inputs.
- Own the result. Revise names, abstractions and docs so the change reads as intentional work. Do not merge code you cannot explain.
Hypothetical example: an engineer asks a coding assistant to add a retry path around a model call. The draft retries every failure, including a malformed request, and can create duplicate downstream work. A taste check narrows retry behavior to the transient cases the service can safely replay, adds a bounded attempt rule, and tests the duplicate path. The useful speed came from generating a candidate; the production design still required human judgment.
Signals that the draft is too expensive
Pause when the suggestion introduces a new package for a small helper, hides state in a global singleton, copies secrets into fixtures, catches every exception, or expands the diff beyond the agreed behavior. Ask for a smaller patch or implement the tricky boundary by hand. Repeatedly regenerating an unclear request usually creates more review work; tighten the acceptance criteria first.
Teams can make this standard practical with small pull requests, CODEOWNERS on sensitive paths, dependency review, and a checklist that fits the repository. Measure whether work reaches review sooner without increasing follow-up fixes, review burden or incidents; do not equate lines generated with engineering throughput.
For team-level guidance on pairing, read pairing with AI without cargo cult. If your project needs a senior engineer to establish the review and delivery loop in a real codebase, see Custom AI Developer services or discuss the current bottleneck.