by Jamon Holmgren on April 14, 2026
The things that have worked the best for me to keep agents like Claude or Codex from writing complete trash code.
(This assumes you're using the top models at medium to high reasoning, and paying $200+/mo for a good plan, not $20.)
- Excellent test suite that the agent has to run and fix if anything is broken, and write its own tests. By far the best way to improve outcomes. Also include linting, type checking, compiling, other static analysis tools and validations, and even access to a debugger if I can make it happen.
- Excellent docs covering systems, code style, testing strategies, and more that I hand-wrote initially and that the agent has to keep up to date with every commit / PR.
-
An opinionated and carefully curated code base with well-named functions/classes/filenames, small files, extremely
flat folder structure, and an
AGENTS.mdthat indexes and describes each concisely. Don't let the intoxicating speed let this get out of hand. You'll pay for it. - Review agents, using codex to review Claude and vice versa. I have Claude spawn codex reviews via CLI and it works super well. Also add in review checklists that it has to use before it's done.
- Well-written specifications that I hand-write and take my time on.
- Review every line of every change that it makes and update docs, tests, or how I write specifications to ensure problems never happen again.
- Run the agents at night so I am forced to improve everything above this one in order to not wake up to slop.
- Be willing to hand-write features and bug fixes from time to time to make sure you stay in tune with the code base.