Skip to content
Open the dashboard
Documentation

What Groundrule does, in one picture

4 min read

Every team has rules about how code should be written. Groundrule puts them in one place, makes sure every coding agent reads them before it writes code, checks every change against them, and shows you what the rules stopped. That is the whole product.

This group of pages explains it in plain words, with examples from a real first week. For the detailed pages, start at What is Groundrule?.

the loopevidence feeds the rulebook; the rulebook feeds everything else
  1. 1One rulebook

    Your team's rules, written once.

  2. 2Agents read it

    AGENTS.md, CLAUDE.md, Cursor, Copilot.

  3. 3Every change is checked

    On the laptop, in the hook, on the pull request.

  4. 4Evidence comes back

    What passed, what was stopped, what to turn on next.

Nothing in the loop is a one-off. Change a rule and the agent files, the checks and the numbers follow.
  1. One rulebook. Your rules live in your workspace, or in a folder in the repository if you use only the CLI. Each rule says what to do, why, and how serious a break is. Most come from ready-made packs; some you write.
  2. Agents read it. One command writes the rules into the files coding agents read first: AGENTS.md, CLAUDE.md, Cursor and Copilot rules. The agent follows the rules while it writes, before anyone checks anything.
  3. Every change is checked. The same rules run as checks: on the developer’s laptop, when the agent finishes a task, and on the pull request. A check names the rule, the file and the line, and says how to fix it.
  4. Evidence comes back. Scans show which rules the code already follows and which it breaks. That is how you decide which rules to switch on, and when.
  • Not a linter. A linter checks syntax and style. Groundrule carries your team’s rules, the ones a reviewer would otherwise repeat in comments, and gives them to agents and checks together.
  • Not an AI reviewer. Checks are deterministic: the same code gives the same result, with no model involved. AI appears only where you ask for it, drafting a rule from a sentence or answering an advisory question on a hosted pull request, and always labelled.
  • Not a gate on day one. New rules start by teaching, not blocking. You switch blocking on rule by rule, once the evidence says it is safe.

Only the CLI. No account. The rules live in the repository, agents read them, the check runs locally and in CI. Everything on these pages up to the workspace works this way.

With a workspace. One rulebook across every repository, a review inbox, rollout stages, the evidence that drives them, and checks on GitHub with nothing in the repository. The pages in this group assume a workspace where it matters, and say so.