jordangrothentic/caliber

Caliber is a small set of Claude Code slash commands, plus a skill you can upload to the Claude website, that tells you which AI model and effort level to use for the task you are about to give Claude, instead of always reaching for the biggest and most expensive option. It only gives advice: it prints the exact commands to run, but it never switches your model or effort setting for you. The core idea is that picking a model should not be about finding the single best one. The strongest model at the highest effort will usually produce the best answer, but it costs more, runs slower, and can even overthink simple tasks. The real question Caliber tries to answer is what is the cheapest model and effort level whose chance of getting the task wrong is still acceptable. If a cheap setup produces a subtle bug, you end up paying twice: once for the run, and again to notice and fix the mistake, so that is not actually the cheap option. When you type a command like slash caliber followed by a description of your task, it looks at how complex the task is, how costly a mistake would be, how much value getting it exactly right adds, and whether a cheaper model could realistically handle it. From that it prints three suggestions labeled Budget, Optimal, and Splurge, each naming a specific model and effort level, along with a confidence rating and a note on what to fall back to if that choice struggles. It also compares its recommendation to whatever model your current session is already using, telling you whether it is a step up, a step down, or the same. To use the full set of five commands you need Claude Code installed locally. To use just the main catch-all command on the Claude website, you need a paid or free plan with Skills and code execution turned on. Installing it is a matter of copying files into Claude's skills and commands folders, either for all your projects or for a single repository.