Vetted AI code is hard to justify

when embedding for real -->

Vetted AI code is hard to justify

Andrew · 17 August 2026 · 3 min read

I have been reflecting on a large optimization I built for my game with the assistance of a frontier coding agent. This optimization took me several days to plan with the agent, about a week to comprehend the massive diff it produced (I review and approve every line), and another week to refactor and refine it until completion.

By the end, I understood everything that had been built—as if I had written it myself—but I was burned out. Untangling and comprehending large amounts of clever, high performance code and deciding if it makes sense within each stratum of abstraction is taxing on the mind.

If I had built it entirely unassisted, I suspect it would have taken me about a month. It would have been difficult, but I would have paid the comprehension costs as I developed it, in manageable chunks, and burned out much more slowly.

I asked Claude make me an interactive visualization for how I perceive these relationships of coding, comprehension, and burnout:

AI & Burnout

Drag each slider to see how different factors can relate burnout.

Model comparing the time and burnout cost of writing code by hand versus reviewing generated code, at varying levels of vetting rigor.

🙂

Vetting rigor?How much of the generated code you actually understand before shipping. 0% is trusting the diff; 100% is knowing it as well as code you wrote yourself.

100%

Output alignment?How close the generated design is to the one you would have written — your decomposition, your conventions. High alignment makes review cheap and rework rare.

40%

Feature size?Scope of the change, 1 to 10. Drives baseline hours superlinearly — a 10 is not ten times a 1, because the pieces interact.

Read : write toll?Burnout per hour of comprehension, against your own authoring at 1.0. Not reading speed — hours are hours. It says rebuilding a model you did not build draws harder on the same day.

2.4×

Hours — manual

Hours — LLM path

Burnout ratio?LLM-path burnout divided by manual burnout. Above 1.0 means the assisted route costs you more than writing it by hand.

Unverified surface?Share of the feature that is both unreviewed and unlike your own design — the part you cannot account for if it breaks at 3am.

Authoring<br>Comprehension<br>Untangling and rework

Manual

LLM-assisted

Bar length is clock time. Colour is what the time is spent doing.

Burnout rises with vetting rigor on the LLM-assisted path and crosses the flat manual baseline.

Reset to defaults<br>A model, not a measurement. Every number here is an assumption you can move.