wtf is understudy?

Understudy is a tool that adds hard rules on top of AI coding agents like Claude Code, so the agent cannot simply say a task is done without proof. It stops the agent from ending its turn until two mechanical checks pass: a log showing that some command actually ran after the last code edit, and a checker that scans the agent's final message for patterns of vague or dishonest sounding language, such as saying "this should work now" without ever running it, or ending with a question that pushes a decision back onto the user. As an example in the README, an agent edits a file, runs nothing, and tries to end by saying the fix is done and asking if the user wants tests added too. Understudy's stop check catches this: no command was run after the edit, and the ending phrasing matches known filler patterns, so the turn is refused and the agent is told exactly what to fix before it can finish. Setup only needs Python 3.9 or newer, using a cross platform install script that can apply the rules to a single project or to every Claude Code project on a machine. There is also a version for Cursor and other agent tools that installs the written guidance without the runtime checks. The project describes its own design as three layers. The first layer is the actual runtime checks that block a turn from ending. The second layer is a set of measurement tools: a standalone linter that can run in CI, plus seven graded practice tasks used to test whether the checks are actually changing agent behavior. The third and smallest layer is a short written document, under a thousand words, that covers judgment calls the mechanical checks cannot catch on their own, such as understanding that a request like "build a client portal" implies parts, like billing and permissions, that were never spelled out. The project is released under the MIT License and has no external dependencies.