sparkline-ventures/product-eval

product-eval is a plugin for AI coding assistants, built to run inside Claude Code or Codex, that helps product managers figure out what to build next and back that decision with real evidence. It connects to the tools where customer problems already show up, such as support tickets in Zendesk, CRM records in Salesforce or HubSpot, issue trackers like Jira, product analytics, and reviews, then pulls out the recurring problems, weighs how strong the evidence for each one is, and ranks them by expected impact. Once a decision is ready, product-eval can hand it off in different forms: a short opportunity brief to align people around the problem, a decision memo that forces a clear yes or no choice with reasoning, a build brief with testable requirements ready to give to a coding or design agent, or a shareable HTML scorecard showing the ranked opportunities. The plugin is designed around a simple chat interface, so a user describes the outcome they want, like ranking opportunities or deciding on one bet, and the tool handles the underlying steps of gathering evidence, scoring it, and checking whether there is enough proof to commit. Decisions go through a two stage check before a recommendation is given. First a sufficiency gate blocks weak cases, such as when evidence is too thin, comes from only one type of source, or the problem itself is described too vaguely. Then a confidence verdict combines the evidence weight, how many different sources agree, any contradictions, and the expected value into a banded result of low, moderate, or high confidence, shown before any exact numbers. The precise scoring numbers themselves are still saved to files for auditing and calibration. All of the investigation work, including business context, discovered sources, evidence, themes, and past decisions, is stored inside a .product-eval folder in the user's own project repository, so it can be versioned with the code and reviewed the same way a pull request would be. Installation options include opening a plugin file directly in Cowork, adding the repository as a plugin marketplace in Claude Code, using a Codex compatible plugin path, or cloning the repository and running an install script that links its 14 skills into the local Claude Code and Codex skill folders.