The case for using AI to write better code, slower

Nolan Lawson pushes back on the dominant narrative that AI coding agents exist to churn out low-quality PRs at maximum speed. He argues LLMs are equally suited to a careful, quality-obsessed workflow, and shares a Claude skill that runs multiple models in parallel — a Claude sub-agent, Codex, and Cursor Bugbot — to surface bugs in a PR, then cross-checks their findings to filter hallucinations. Using several different models dramatically cuts the false-positive rate.

The technique reliably uncovers far more issues than a developer can reasonably fix, spanning critical security flaws down to misleading comments. Lawson’s workflow tackles criticals and highs first, skips fixes where the cost outweighs the benefit, and abandons PRs entirely when too many serious issues suggest the underlying approach is wrong. The review often exposes pre-existing bugs, sending him on tangential cleanups rather than boosting raw velocity.

The payoff isn’t 10x productivity but better codebase health and deeper familiarity with a system’s failure modes — historically how good engineers learn unfamiliar code anyway. Lawson frames it as an LLM-amplified version of methodical engineering, aimed at developers currently shipping multi-hundred-line PRs they don’t fully understand.