markusnigit/human-uat-kit

Human UAT Kit is a set of written prompts that turns an AI coding agent into a simulated testing department for a game or piece of commercial software before it ships. Instead of only checking that code matches its written specification, the kit is built around a different problem: parts of a product can each work correctly on their own but still feel wrong together, or quietly fail the user in ways a spec never anticipated. The kit's author found that a short human testing session once caught defects that a deep automated code review had missed, and the kit tries to catch that category of problem on purpose. The kit works by giving the AI agent two separate roles. An Operator reads everything about the project, including internal documents and code, and sets up the test run. A Tester only receives what a real user of the product would know, and before looking at any internal documentation it writes down what a reasonable user would expect from a product like this. When those expectations disagree with what the documentation says the product should do, that disagreement itself becomes a finding worth reporting. The run proceeds through phases covering code health, hostile input handling, hundreds of simulated user sessions, slower think aloud style testing of the experience itself, a panel of simulated personas representing different kinds of users, and a final release readiness verdict with a ranked list of what still needs a human to check by hand. Anything involving actual visuals, touch, or sound is marked as a low confidence guess rather than a real finding, since the agent cannot see or hear the product directly. Using the kit means copying a handful of markdown files into a project, filling in a profile describing how the AI can operate the product without a screen, such as through an API or command line, and then telling the agent to run the main prompt file. It was built and tested using Claude Code, though the prompt itself does not require any particular agent. The kit explicitly refuses to test systems the user does not own or touch production data and real payments.