What they did
Claire builds systems with Claude Code and kept finding flaws only after running them. On 2026-08-27 a job-history marker built from the workspace folder path had to be redone because repeated failures in one folder all got the same marker. That day Claude Code built a shell script, soudan, that it calls through Bash to ask Codex for a design opinion. The script runs Codex on a ChatGPT subscription and refuses to run if an API key is set, caps each consultation at three round trips, cuts short questions off at 90 seconds and writes the reply to a file. Codex has an AGENTS.md telling it to give one judgment, admit when it does not know and check seven angles, starting with whether names or paths are used as identifiers. Every new build now gets a design review before coding and a code review after a separate agent has run it.
How it went
Sent the design that had just been rebuilt, Codex replied in three lines that folder paths make bad markers because moves and reused paths confuse jobs, which was the exact failure. A design for posting three AI-made videos a day drew over 30 points with sources, mostly on law and platform terms. One of them, that unaudited API projects get uploads forced to private, did not match reality: 16 videos were already public. The post-build review caught two flaws the run check had missed. Reply times ranged from 3 seconds to over 5 minutes.
Worth knowing
A second agent adds opinions, not facts; check its claims before acting on them. Claire limits self-initiated consultations to 10 a day, after which Claude Code must ask a person, but never caps the reviews built into the process. Showing the decisions rather than the code worked best.