ralfyishere/rules-with-receipts

Rules with Receipts is an installable set of behavior rules and skills for Claude Code, the AI coding agent tool, designed to make agents more disciplined rather than smarter. The project's core argument is that most agent rules files promise big improvements with no proof behind them, while agents keep failing in predictable ways, such as claiming a task is finished without actually testing it, or making unrequested changes alongside what was asked. This project tries to fix that by shipping alongside its own testing system that measures whether the rules actually change behavior, and by publishing the results honestly, including results that do not flatter the project. You install it by running a bootstrap script against your own project folder, or by cloning the repository and running an install script directly if you would rather review the code first. After installing, it adds 37 skills, 14 rules that stay active at all times, and gates that block actions like pushing code or publishing a release until a security scan passes, even if the agent forgets to check on its own. The published evaluation results are notably candid. The author found that a plain, unmodified model already passes most of their own tests without any rules installed, and that the biggest measurable improvement from installing this pack is in staying within the requested scope of a task rather than wandering into related changes. An earlier claim about a specific bonus effect, catching an adjacent bug more reliably, did not hold up in later, larger test runs, and the author documents that correction openly rather than hiding it. An independent third party test also did not fully reproduce the original headline result. The author is explicit that this project is not an intelligence upgrade, not a general capability benchmark, not a safety guarantee, and not a substitute for a human reviewing the agent's work. It is released under the MIT license.