generate-workbench

generate-workbench

Bootstrap step 2. Turn a .persona/ draft into a real, minimal workbench repo.

The mechanical part — laying the scaffold and composing AGENTS.md from the framework core + the persona overlay — is deterministic and identical every time, so it lives in a bundled script: scripts/compose.sh. Your job in this skill is the judgment around it: get the target repo, locate the framework core, run the script, then clean up and hand off. Don't re-derive the scaffolding by hand — call the script.

Input = .persona/ (the contract from interview)

• .persona/overlay.md — filled AGENTS.overlay.md (the 7 persona cells)
• .persona/meta.yaml — repo: owner/name or path: …, plus language:
• .persona/codebases.yaml — (optional) initial work targets

If .persona/ is absent, run interview-for-personalizing first — don't invent a persona.

Procedure

1. Obtain the target repo from .persona/meta.yaml.

   • repo: owner/name → create it if needed (gh repo create owner/name --private), then clone, or use an existing local clone. Confirm with the user before creating anything remote — repo creation is outward-facing.
   • path: … → use that local directory (init git if empty).

2. Locate the inputs — both bundled in this plugin (so it works when installed). scaffold/ (with AGENTS.core.md, AGENTS.overlay.md, templates, empty docs, codebases.yaml) ships inside this bootstrap plugin at ${CLAUDE_PLUGIN_ROOT}/scaffold/. No cross-plugin lookup is needed.

3. Run the compose script:

   "${CLAUDE_PLUGIN_ROOT}/skills/generate-workbench/scripts/compose.sh" \
     --persona  <.persona dir> \
     --core     "${CLAUDE_PLUGIN_ROOT}/scaffold/AGENTS.core.md" \
     --scaffold "${CLAUDE_PLUGIN_ROOT}/scaffold" \
     --out      <target repo>
   

   It lays templates/ .github/ docs/(empty) codebases.yaml, writes AGENTS.overlay.md from the persona, and composes AGENTS.md = core + overlay (inline). It does not copy the engine — that's the plugin's job.

   The script self-checks: it rejects an empty/half-filled .persona/ up front, and after composing it asserts post-conditions (AGENTS.md has both core+overlay markers, overlay/docs/codebases present, engine not copied, CLAUDE.md symlink) — failing loudly. So you don't eyeball the result; if compose.sh exits 0, the workbench is well-formed. If it exits non-zero, read the FAIL lines and fix the input.

4. Discard .persona/ — it's temporary scratch; its content now lives in the repo.

5. Commit and hand off. Make the initial commit (in the user's output language per the persona), then point the user to their first task: workbench:task-start <issue> (engine plugin). Remind them the workbench plugin must be installed/enabled for the engine commands to work.

Output (what the user gets — minimal by design)

my-workbench/
  AGENTS.md          composed: framework core + your persona (do not edit directly)
  AGENTS.overlay.md  your rules — edit here, then recompose
  codebases.yaml     your targets (or empty template)
  docs/              empty knowledge skeleton (index + empty anchor tables + log)
  templates/  .github/   profile defaults you can tune

No utils/, skills/, or bin/ — the engine is the installed workbench plugin. That's the whole point: your repo carries only your stuff.

Boundaries

• Don't copy the engine into the user repo — keep it minimal; the plugin provides it.
• Don't create remote repos without confirming (outward-facing).
• Recompose (re-run from edited AGENTS.overlay.md) is how rules change later, and how framework upgrades flow in (the script re-reads the plugin's new core).