Using AI to write better code more slowly | Read the Tea Leaves

The author argues AI can be used to write higher-quality code more slowly, not just to produce fast low-quality output.
LLMs and agents are described as very effective at finding bugs in a codebase, including subtle bugs in recent Anthropic and OpenAI models.
The main difficulty is not discovering bugs but prioritizing them and validating which ones are real.
To reduce hallucinations and false positives, the author combines multiple models on the same PR review.
The workflow uses a Claude sub-agent, Codex, and Cursor Bugbot to find and rank bugs as critical, high, medium, or low.
After the models finish, the human reviewer checks the findings, does additional research, and writes the final report.
The author’s definition of a bug includes violations of KISS and DRY, inaccessible HTML/JSX, and missing SQL indexes.
The practical result is often fixing critical and high-priority issues first, while skipping lower-value fixes when appropriate.
If the review reveals the underlying approach is flawed, the author may abandon the PR altogether.
This slower AI-assisted process may not boost raw productivity, but it improves code quality, codebase health, and understanding of failure modes.

Your notes

Save this item to your library to add private notes.