global-plugin

global-plugin

Source repository for the global-plugin Claude Code plugin — company-wide guardrails for full-stack React/Next.js + Node/NestJS + Prisma/Postgres + AWS projects, with optional React Native mobile.

The shipped plugin lives entirely under plugin/; see plugin/README.md for the consumer-facing description, full skill catalog, and install instructions. Everything else in this repo (docs/, scripts/, .husky/, this README, CLAUDE.md) is dev infrastructure for building the plugin and never reaches a consumer.

What the plugin does

Ships ~24 skill-format guardrails ("guards") that activate when an agent works on a consumer project, organized by concern:

• Architecture & structure — monorepo dependency boundaries, Next.js app structure, NestJS service boundaries, frontend implementation, mobile implementation
• Data — Prisma access patterns, state integrity
• Integration & async — contract safety, queue/retry, resilience
• Security & config — auth/permissions, secrets, supply chain
• Quality — TypeScript rigor, test strategy, coverage gaps
• Frontend quality — accessibility, performance budget
• Ops & risk — change-risk evaluation (covers blast radius, rollback path, deploy strategy, monitoring, stakeholders), infra change, AWS deploy, CI/CD, observability-first debugging

Each skill is self-contained — domain skills hold their own rules. Two lightweight hooks ship: a UserPromptSubmit hook injects a brief skill-loading-discipline reminder, and a Stop hook deterministically runs the project's lint / typecheck npm scripts when files have changed during the session, blocking the stop on failure. No MCP servers and no loggers ship with the plugin.

The full mechanism inventory — exact skill list, hook config, recommended companion plugins — lives in plugin/README.md.

Repository layout

Path	Purpose	Shipped to consumers
plugin/	Plugin runtime — manifest, skills, hooks, onboarding scripts	Yes
docs/	Design notes, plans, specs, audits, workflows	No
docs/superpowers/{plans,specs,workflows,audits}/	Structured maintainer artifacts (dated YYYY-MM-DD-<topic>.md)	No
scripts/	Skill-verifier (TypeScript) + vitest harness	No
.husky/	Pre-commit hook — runs pnpm verify on staged SKILL.md	No
package.json, pnpm-lock.yaml	Dev deps only (vitest, husky, tsx)	No
CLAUDE.md	Maintainer-mode project instructions	No

Local test

From any project directory:

claude --plugin-dir /absolute/path/to/global-plugin/plugin

Inside Claude Code:

• /help lists global-plugin skills.
• /mcp is empty (the plugin ships no MCP servers; configure your own in your project's .mcp.json).
• /global-plugin:architecture-guard triggers a skill.

To exercise the plugin the way a consumer would (without this repo's CLAUDE.md polluting the session), run the command above from a clean directory — never from this repo root.

Maintainer workflow

pnpm install                                  # one-time
pnpm test                                     # vitest suite (skill-verifier)
pnpm verify plugin/skills/<name>/SKILL.md     # one skill

The husky pre-commit hook runs pnpm verify automatically on staged plugin/skills/*/SKILL.md files. The verifier itself is plain TypeScript under scripts/verify/ (parser, runner, checks) with a vitest test suite — no plugin primitives, no AI in the loop.

Consumer-facing docs

See plugin/README.md for the consumer-facing skill catalog, hook list, MCP setup, and install instructions. (A one-command new-project setup script is being reworked — see plugin/README.md's deferral note.)