Agent-Skills.md

Agent Skills: Self-Install

>

UncategorizedID: mattgierhart/PRD-driven-context-engineering/ghm-self-install

Repository

mattgierhart/PRD-driven-context-engineering
mattgierhartLicense: MIT
2069

Install this agent skill to your local

pnpm dlx add-skill https://github.com/mattgierhart/PRD-driven-context-engineering/tree/HEAD/.claude/skills/ghm-self-install

Skill Files

Browse the full folder contents for ghm-self-install.

Download Skill

Loading file tree…

.claude/skills/ghm-self-install/SKILL.md

Skill Metadata

Name
ghm-self-install
Description
>

Self-Install

Install the methodology into a target repo from the install manifest, the way Xantham's blueprint installs from markdown: the agent runtime is the installer, so the whole thing runs inside your Claude Code subscription — no API key, no metered calls, no service to stand up.

Two paths, one manifest. This skill is the interactive path. install.sh is the deterministic CLI path. Both read .claude/install-manifest.yaml, so they can't drift. Prefer driving install.sh from this skill; fall back to manual copies only when the script isn't reachable.

Consumes

  • .claude/install-manifest.yaml — authoritative framework / template_seed / never_touch lists (the keystone; never hardcode file lists).
  • install.sh — the deterministic installer this skill normally drives.
  • Safety rules from ghm-template-sync Phase "Safety Rules" — never overwrite MEMORY.md, SoT/*, PRD.md, README.md content.

Produces

  • An installed framework in the target repo (no new SoT IDs — this skill places the engine, it doesn't author specs).
  • A verification report (hooks emit valid JSON; readiness.py runs).

Workflow

Phase 1 — Preflight & mode detect

  1. Confirm git, python3, awk are present; warn if the target isn't a git repo.
  2. Detect mode: greenfield (no .claude/) vs brownfield (existing .claude/). Brownfield means merge, never overwrite product.
  3. State the trade explicitly: this installs files only — it costs subscription tokens for the wizard turn, zero API.

Phase 2 — Wizard questions

Ask only what changes the outcome (honor the execution mode's budget):

  • Target directory (default: current repo).
  • Domain profile: product (default) · library · infrastructure · research.
  • Subset [standard+]: keep all 4 agents + all skills (default), or trim to a role subset. Quick mode skips this and installs everything.

Phase 3 — Install (drive install.sh)

Run the deterministic installer so behavior matches the CLI path exactly:

bash install.sh --target <DIR> --profile <PROFILE>        # add --dry-run to preview
  • Show the --dry-run plan first [standard+], then execute.
  • If a framework file shows drift, surface the diff and only re-run with --force after the user confirms (mirrors ghm-template-sync "show diff before destructive").
  • Never write anything in the manifest's never_touch list.

Phase 4 — Verify (trust-but-verify, the Xantham verify-blueprint.sh analog)

  1. Run each .claude/hooks/*.sh in the target; assert valid JSON on stdout.
  2. Run python scripts/readiness.py run — a non-zero exit on an empty scaffold is the expected BLOCK gate (no content yet), not a failure. Report the score.
  3. Print next steps: customize README.md + PRD.md, then "Let's frame the problem".

Anti-patterns

| Pattern | Fix | |---------|-----| | Hardcoding the file list in the skill | Read .claude/install-manifest.yaml | | Overwriting a brownfield repo's PRD.md/SoT | Honor never_touch; seed templates once | | Treating readiness BLOCK on a fresh scaffold as a bug | It's the gate working — report the score | | Re-implementing copy logic instead of driving install.sh | One code path = no drift |