Getting good at AI takes time and practice

Sean Goedecke, a software engineer, argues that working well with AI agents takes practice, and that reviewing AI-written code needs habits of its own.

Sean Goedecke writes essays about software engineering practice. In this one he answers engineers who say that using AI needs no skill and that any tricks will expire with the next AI model. He names four skills that he says last across AI model generations, with examples from his own work.

• Skilled users give an AI agent the missing context and the reason for the task, so that it can make decisions instead of guessing.
• Goedecke asks agents for ideas on how to structure a service, but not for new features. He also asks them to test their own changes locally.
• In his account, AI-written code rarely has counting errors. It often adds machinery for unimportant cases, or trades a real requirement for an invented one.
• He warns that fluent, heavily commented and fully tested AI code misleads reviewers. He tells them to distrust additions such as a new lock on shared data.

An engineering team can teach the review of AI-written code as a separate skill, with its own list of common faults. This is an opinion essay based on the author's experience, without measured results.

Derived by Working Surface from the article; more in the Playbook. Source line: Reviewing AI-generated code is a substantially different skillset from reviewing human-written code.

29 October 2026: as AI agents raise the volume of work to check, teams move the check closer to the work: fixed rules for what agents may fix, design rules written where agents read them, reviewers trained on how AI-written code goes wrong, and leaders who do the work and are reviewed in public.