Case study: this website
The site you are reading runs on the methodology it documents. It dogfoods Deep Work Plan: the same skill, the same /init flow, the same .dwp/ plans any other repository would use. This is a real account, not a hypothetical.
Before
Like most repositories, this one was not built for agents. Context lived in people’s heads and scattered notes, there was no single source of truth an agent could read, and onboarding a new agent meant re-explaining the project every session. Long-horizon work drifted.
Adopting DWP
The repository was made AI-first with a single Deep Work Plan, decomposed into atomic tasks with validation gates:
- Install the Deep Work Plan skill as a reference install, pinned by
skills-lock.json. - Run onboarding to generate a reasoned
AGENTS.md, thedocs/tree, and per-module docs from the repository’s real stack. - Build the cross-agent
.agents/kit — thindwp-*command delegators and a catalog that matches what is on disk. - Scaffold the gitignored
.dwp/workspace for plans and drafts. - Verify conformance with
/dwp-verify.
Each task validated against the repository’s real gates — biome, astro:check, the test suite, the production build, and the agent-endpoint parity check — before it was marked complete.
After
The repository is now AI-first by its own specification: a reasoned AGENTS.md, a categorized docs/ tree, a cross-agent .agents/ kit, and a gitignored .dwp/ workspace. Any agent — Claude Code, Cursor, Codex, Gemini, Copilot — can open it, read the harness, and execute long-horizon plans without per-session hand-holding.
Outcome
The methodology proves itself on its own source: this site is built and maintained the same way it tells you to build yours — by following /init.md. If the standard works here, in production, it works for your repository too.