Every AI coding agent reads an instruction file — AGENTS.md, CLAUDE.md, or a system prompt. Most are copy-pasted and half-stale. Groundrules is a visual builder for them: drag battle-tested blocks (conventions, workflows, hard rules with verification probes) onto a canvas, compile to any target your tools read, then Test Drive against 12+ live models (Claude, GPT, Gemini, Grok…) to see whether the agent actually obeys — before you commit the file. Core builder is free.

Hi Product Hunt 👋 I built Groundrules because I kept rewriting the same instruction files for every project — and because an agent once "fixed" my failing tests by deleting them. ✓ 0 tests failing. Technically correct. My fault, really: nobody had written the rules down. So: a visual builder for the files agents actually read. Assemble a config from ~80 battle-tested blocks (conventions, workflows, hard rules like "never commit secrets" — with probes to verify they hold), and compile it to AGENTS.md, CLAUDE.md, a plain system prompt, or settings.json + ESLint rails. The part I'm proudest of is Test Drive: run your rules against a dozen live models and see whether the agent actually obeys them. A rule that doesn't survive contact with the model is just a wish. The core loop — building, exporting, sharing — is free and stays that way. Pro ($6/mo) covers what costs me money: hosted test runs, version history. Honest rough edges: the community catalog is young (submissions welcome — review is currently just me), and Team features are a waitlist. I'd love to know: what rules does your team enforce that belong in the catalog? The best blocks so far came from other people's war stories.

honestly the test drive against 12+ models thing is genuinely clever, like actually seeing whether the agent obeys before you ship the file saves so much guesswork. solid execution.

The test-drive-against-live-models idea is genuinely clever, turning what used to be blind faith in a prompt file into something you can actually verify before shipping it.

honestly the test drive against 12+ models thing is genuinely clever, like actually seeing whether the agent obeys before you ship the file saves so much guesswork. solid execution.

The test-drive-against-live-models idea is genuinely clever, turning what used to be blind faith in a prompt file into something you can actually verify before shipping it.

Groundrules

Assemble the ground rules your AI coding agents follow

Assemble the ground rules your AI coding agents follow