v1.1 · Groundwork MIT · works in 50+ harnesses

Context is the new code

Most teams write it once and never touch it. We make it shippable: versioned, reviewed, wired to the checks that prove it works.

The ten commands Install guide

What you get

One command. Real files.

groundwork init reads your repo, asks five questions, and writes the files every other command assumes exist.

One canonical AGENTS.md. A 5-line pointer per harness. ADR scaffolding. A pre-commit hook that fails when context drifts from the code. No TODO: implement. No 800-line CLAUDE.md.

How init works

$ groundwork init

your-repo/
├── AGENTS.md canonical, ~70 lines
├── CLAUDE.md → AGENTS.md symlink
├── .cursor/rules/main.mdc 5-line pointer
├── .github/copilot-instructions.md 5-line pointer
├── .windsurf/rules/main.md 5-line pointer
├── .context/hooks/check-context.sh pre-commit drift
└── docs/decisions/
├── README.md
└── negative-space.md rejected paths

The skill

Ten commands. Start with three.

All commands

init

Bootstrap a repo with the full context engineering scaffolding.

A new repo, or an existing repo with no CLAUDE.md / AGENTS.md / .cursor/rules yet.

verify

Bake test, lint, and typecheck commands into context so the agent self-corrects.

Every repo. Every time. The single highest-leverage move in this skill.

audit

Score the repo across the five context layers and name the gaps.

"How good is our AI setup?" Or before a migration, a new-hire, or a handoff.

The foundation

The model underneath.

All seven pages

The five-layer stack

Every piece of context belongs in exactly one layer. Mixing them is the root cause of "the agent did the right thing locally and broke everything else".