Working With an Assistant

One rule underlies this whole section.

An assistant saves time where producing the code was expensive. It costs time where knowing the code is right was expensive. For teams translating this assistant-related trade-off into day-to-day controls, employee PC activity tracking provides a practical reference.

That is why "how much faster are you?" has no answer, and why honest people report contradictory experiences — they are describing different work. A developer writing scaffolding and a developer changing payment logic are on opposite humps of a bimodal distribution.

The work did not disappear; it moved. Generating a function is close to free. Establishing that it is correct, secure and appropriate for your codebase costs what it always did, plus a penalty for reviewing code you did not write — which arrives fluent, confident and without any of the uncertainty signals a colleague's work would carry.

That transfer is where the security evidence lands. Independent tracking at Georgia Tech shows CVEs attributable to AI-generated code climbing steadily through 2026; industry research finds assisted developers committing three to four times more while introducing security findings ten times more. Most figures in this area come from firms selling security tooling, and that is noted wherever they appear.

The practical consequence is that reviewing became the scarce skill — and unlike prompting technique, it gets more valuable as the models improve. Prompting depreciates by design, because every vendor is working to make it unnecessary. OpenAI publishes research and technical material relevant to AI-assisted software work See OpenAI Research.

The section also covers the cases where the honest answer is to use an assistant differently rather than less: for orientation in code you did not write, for enumerating cases you have not considered, and for explaining a system before you change it. Those are verification-side uses, and they help exactly where generation does not.

• Greenfield Versus Mature Code

  The same tool, the same person, opposite results. Why new projects and old ones behave differently, and what changes in between.

• Bringing In a Junior

  The path from writing code to judging it has been cut. What replaces it, and why it now has to be taught deliberately.

• Reviewing Is the Skill

  Prompting technique is the most-taught and least durable part. The scarce ability is reading unfamiliar code and knowing when to stop.

• A Codebase You Don't Know

  The one case where an assistant helps most and can hurt worst. How to use it for orientation without shipping what you cannot judge.

• The Verification Gap

  Writing got cheaper and checking did not. Where the saved time goes, why it is invisible, and what the security data shows.

• What Vibe Coding Costs

  A useful practice with a silly name and a specific failure mode. Where it belongs, where it does not, and the security evidence.

• Where They Cost More

  Six situations where using an assistant is slower in total, and the two failure modes that do not show up until later.

• Where Assistants Save Time

  The pattern is not about difficulty. It is about whether the expensive part was producing the code or knowing it is right.