Thoughts is a set of instructions that tells a model to examine ideas with evidence instead of agreement. I tested it against no instructions and a one-line prompt on four models: it helped Gemini and Sonnet, on GPT it was no better than the one-liner, and on Claude Opus 4.6 the result depended on the grader. Below are the conversations that led me to build it, then the full results, including the failures, with every answer and grade in the test kit.
Read in full here: