Groundwork — Development skill for Claude Code

groundwork

A Claude Code skill that gives a project a memory.

You answer questions once. It writes the notes. Every session after that already knows your project, your rules, and where you left off.

The problem

Open a new chat and the agent knows nothing. It doesn't know what you're building, that you hate four-level nesting, or that you already tried Redis and dropped it. So you explain. Next week you explain again.

Groundwork writes all of that down once, in files the agent reads on its own.

How it runs

flowchart TD
    A["/groundwork"] --> B{Any code here?}
    B -->|Yes| C[Read the repo first]
    B -->|No| D[Nothing to infer]
    C --> E{External brief?}
    D --> E
    E -->|Yes| F["Fetch it, extract requirements,
read them back to confirm"]
    E -->|No| G[Interview, five rounds]
    F --> G
    G --> H["Write CLAUDE.md + docs/"]
    H --> I[Humanize what people will read]
    I --> J[Offer to block dangerous git commands]

What it writes

CLAUDE.md              read every session, capped around 60 lines
docs/
  brief.md             only when there's an external brief
  overview.md          what this is and who it's for
  architecture.md      how it's built, and what you rejected
  standards.md         taste, testing, prose voice, security
  progress.md          where you are now
  changes.md           append-only decision log
  handoffs/            session notes, one file per stop

Only `CLAUDE.md` loads every session, which is why it stays small. It holds your rules and an index. The rest gets opened when it's relevant and costs nothing the rest of the time.

The interview

Five rounds, three questions at a time, about fifteen minutes.

It reads your repo before asking anything, so it confirms rather than interrogates. It won't ask what test framework you use when `package.json` already says. Say `skip` to any round and it writes `Not yet defined` instead of inventing an answer.

It also asks what you *dislike