wtf is bakeoff?

Bakeoff is a skill for Claude Code that helps you decide between several possible solutions to a problem by turning the decision into a small, judged competition. Instead of you having to figure out what criteria matter and score each option yourself, you describe the decision in plain language, and the tool generates a handful of different candidate answers, works out what should be scored and how much each factor should count, and then judges every candidate against that scoring sheet. The judging step uses two separate judges that each see the candidates in a different random order, and their scores are then reconciled by a script rather than by a debate between them, since the authors found that debating tends to amplify bias instead of removing it. Once a leading candidate is chosen, the tool runs an extra check that actively tries to argue against that winner, so a choice that only looks good on the surface has a chance to be caught before it is recommended. When the top two candidates are strong in different ways, bakeoff can combine the best part of the runner up into the winning answer, but only keeps that combination if it scores better than the original winner on its own. For decisions that depend on real facts, like a library version or details of your own codebase, the tool reads or searches for that information first, and marks anything it could not verify as coming from its training knowledge only rather than a checked fact. The tool is meant for situations where there are several reasonable options, getting it wrong would be costly to undo, and it is not obvious what should count as a better choice, such as picking an architecture, a library, or a migration approach. It is not meant for questions that a test or a type check could already answer, or for judging a single option with nothing to compare it against. It installs as a self-contained skill for Claude Code, through a skills command line tool, a Claude Code plugin marketplace, or by cloning it directly into your skills folder, and it needs Node.js available on your system since a scoring script runs under Node. Every run saves a written report with the recommendation, the full scores, and the reasoning, so the decision can be reviewed again later.