Candor is my free, MIT-licensed Claude Code plugin: twelve work-mode "personas" that make an AI assistant direct, verification-driven, and honest instead of agreeable. A debugging persona that demands a reproduction before guessing. A data persona that reads numbers skeptically. A decision persona willing to say no. Each one is documented, benchmarked with eval scripts, and installable in minutes. I built it because of what I learned running AI workflows on a real business (my mom's Shopify boutique, the Krafty Bandit project below): the failure mode of AI at work is not wrong answers, it is confident agreement. The workflows that shipped 2,861 catalog fixes in one weekend only worked because every call got checked by a human, and Candor bakes that discipline into the assistant itself. The repo includes twelve skill definitions, evals with published benchmark results, full documentation, and contribution guidelines. It is labeled experimental (v0.5.x) and actively maintained. Free forever. It is also the companion tool for my YouTube channel, where I show the personas doing real work on a live store. github.com/itsryanlenk/candor