devdacian/touchstone

Touchstone is a process, not a running app, that helps you improve the instructions (called skills) you give to AI coding assistants like Claude Code or Codex. You give it an idea, a new feature request, or a bug report, and it automates a full cycle: gathering context, drafting an initial approach, getting a second opinion from a fresh AI session, writing a plan to a file, running review rounds on that plan, implementing the change, then running review rounds on the implementation too. The review step is the core idea. Each round hands a reviewer a prompt naming specific things to look for, like stale references or sequencing bugs, rather than a vague request for feedback. Every issue the reviewer raises is treated as a claim that needs checking, such as looking at the actual line of code or running the script in question, before the plan owner decides to accept, reject, change, or set it aside. When a second AI model from a different company is available, it reviews the same plan and code in parallel, since a different model tends to catch different kinds of problems. Installing it means copying one folder into your project and adding a short pointer to your own project instructions file. It does not overwrite your existing setup, and by default your project keeps the Touchstone methodology files while ignoring files it generates while running. There is no server to run and no extra software required beyond the AI coding tools you already use. If the second AI model is not reachable due to an authentication or quota problem, the process still finishes using only the first model, with a clear note that this happened. The project is aimed at teams and individuals already using AI coding assistants who want their custom skill instructions to improve in a structured, repeatable way instead of by trial and error. It is released under the MIT license.