What is reflect?

Reflect is a Claude Code skill built to fix a specific problem people notice with AI assistants, they tend to agree with whatever direction you lean. Ask if you should launch something and it finds reasons to launch, ask if launching is a bad idea and it finds reasons against, even though nothing about the actual decision changed. Reflect is meant to counter that by making the AI argue against its own first answer before it shows it to you, rather than simply agreeing with your framing. You use it by typing a single sentence, something like reflect on whether I should rewrite this module or keep patching it. After running, it gives you a direct answer stated plainly rather than a list of vague considerations, along with a rough confidence number so you know whether to act right away or test the idea first. It also tells you the specific next step to take, and lists the one to three things that would change its answer, so you are not left relitigating the same decision later. What makes it different from generic advice is that it actually looks at your project before answering. It can run git log to understand why a piece of code exists instead of guessing, it looks for real base rates rather than asserting made up statistics, and if the answer depends on something checkable like a quick test or existing usage numbers, it actually checks that first instead of just describing what checking would involve. The underlying process draws on the decision making approaches of well known thinkers like Jeff Bezos, Charlie Munger, and others associated with Farnam Street, compacted into a set of 23 mental models. It first asks whether a decision is easy to undo, since reversible decisions get a faster, less rigorous answer, while harder to reverse ones get the full deeper analysis. The tool also keeps a running score of how your past predictions with reflect turned out, so you can see over time how well calibrated your instincts have been.