Four models checking each other
I stopped asking one model to check my work and started making four of them argue about it. It won a hackathon in August, so I open sourced it.
At some point I stopped asking one model to check my work. I started sending the same thing to four of them, separately, and making them argue about it.
It worked better than anything else I had tried, so in August I took it to the Formidable Builders Skillathon, Europe’s first agent skill hackathon, supported by OpenAI’s Codex and Apify, and built the entry that way. It won first place out of the 34 that shipped that evening, in a full house of 69 builders.
I have been using the same technique every day since, on everything I build, Anima Felix included. So I cleaned it up and made it public: decent-skills, MIT. Agent skills I built for my own use and decided to open source.
/council-review is that technique in one command. Three models get the same work to review, none of them able to look at the repo, none of them seeing what the others said. The fourth is whichever one I happen to be working in: it collects what comes back, checks every claim against the actual files, and throws out the ones that do not survive. Nothing changes until I say so.
The isolation is the whole point. A model that sat next to you while you wrote the code agrees with you too easily. Cut it off from the repo, from your conversation, from the other reviewers, and what you get back is an opinion instead of an echo.
It is not magic, and the honest failure is the interesting part. In one run three of the four independently flagged the same thing, and all three were wrong in the same way, because they had all been handed the same limited context. That is exactly why something has to check them against reality before any of it reaches you.
The other two skills are smaller. /wawa tells me where a project actually stands, by reading the repo first and my own notes second, because notes are just a guess with a date on them. /init-repo-docs writes the status, todo and decision files from the git history, so a project remembers itself.
In Claude Code:
claude plugin marketplace add cochinescu/decent-skills
claude plugin install decent-skills@decent-skillsThen type /council-review, /wawa or /init-repo-docs. For Codex and Grok, clone the repo and run ./install.sh. The README has the rest, including which CLIs you need for a full review.
One thing worth saying plainly: a review sends the code you selected to OpenAI, Google and xAI, through your own accounts. That is the trade. Four opinions from four companies, and none of them is the one that wrote it.
