parvez

claude-skills

⚠️ Work in progress. An evolving, open-sourced toolkit of Claude Code building blocks - skills, hooks, and commands. Some are battle-tested, some are early. Expect churn. Reuse the patterns at your own discretion and adapt to your stack.

Not just skills - a toolkit, because the right Claude Code primitive depends on the job:

Primitive	Use it for	Triggered by
Skill	knowledge / a workflow the model applies with judgment ("how to do X")	the model, on a matching task
Hook	deterministic must-run-on-event enforcement ("always do X when Y")	the harness, automatically
Command	a saved prompt the user fires	/name

Rule of thumb: "always do X when Y happens" is a hook, not a skill - the model forgets, the harness doesn't. "how should I approach X" is a skill.

Skills (skills/)

Skill	What it does	Scope
vps-deploy	Vendor-neutral push-to-deploy CI/CD to a VPS (any runtime: Node/Python/Go/.NET), Caddy auto-TLS, OAuth/OIDC redirect setup, DNS (proxied-by-default), and baseline hardening (fail2ban, SSH-port rotation, firewall, CDN origin lockdown).	generic
honest-health-check	Design health endpoints that assert real readiness (DB/deps reachable) instead of a hardcoded 200, and wire them into deploy gates + container/orchestrator probes. Pairs with vps-deploy.	generic
clean-architecture	Rulebook + bootstrap + config-driven guard for Clean / Hexagonal architecture in any language (.NET/TS/Java/Go/Python). The Dependency Rule, ports & adapters, per-stack layout, and a layering check to enforce boundaries.	generic

Hooks (hooks/)

Hook	What it does	Install
secret-hygiene	Block a commit that introduces a secret (gitleaks when present, conservative regex fallback).	git pre-commit hook, or a Claude Code PreToolUse hook - see hooks/secret-hygiene/README.md
safe-commit	commit-msg validator: no stray @, conventional subject, length; optional em-dash + AI-coauthor rejection.	git commit-msg hook - see hooks/safe-commit/README.md

Commands (commands/)

User-fired prompt shortcuts. None yet - coming as patterns stabilize.

Install (any machine)

This repo is a Claude Code plugin marketplace. The skills ship as one plugin, parvez, so on any machine:

/plugin marketplace add parvez-ahammed/claude-skills
/plugin install parvez@parvez-tools

Every skill goes live, namespaced parvez:<skill> (e.g. parvez:vps-deploy, parvez:clean-architecture, parvez:honest-health-check), triggering on its description. Update later with /plugin marketplace update parvez-tools.

Local authoring loop (test before pushing): /plugin marketplace add . from a clone, then install — relative paths resolve against the repo.

Layout

claude-skills/
  README.md                  - this file
  LICENSE                    - MIT
  INDEX.md                   - catalog
  .claude-plugin/
    marketplace.json         - marketplace manifest (name: parvez-tools)
  plugins/
    parvez/                  - the installable plugin (namespace prefix)
      .claude-plugin/plugin.json
      skills/                - skills (one folder each)
  _template/                 - starting point for a new skill
  hooks/                     - git hooks (script + install notes per hook)
  commands/                  - user-fired commands
  sync-skills.ps1            - legacy: copy skills into ~/.claude/skills (dev only)

Use

Skills - install via the plugin (above). They trigger on their description. To add a skill: drop a new folder under plugins/parvez/skills/, commit, push; installed users pick it up on /plugin marketplace update.

Hooks - these are git hooks (commit-msg / pre-commit), not Claude lifecycle hooks, so install per-hook (a git hook, or a settings.json snippet). See each hook's README.md. The update-config skill can wire Claude Code hooks into settings.json.

sync-skills.ps1 is the legacy path (copies into ~/.claude/skills on this box only) - kept for fast local iteration; the plugin marketplace is how it goes cross-machine now.

Writing a good skill

• Pushy, specific description - the only trigger signal; name the contexts and phrases it should fire on.
• Keep SKILL.md lean; push detail into references/ (loaded on demand).
• Bundle deterministic steps as scripts/.
• Explain the why; avoid rigid ALWAYS/NEVER walls.
• If it's really "enforce X every time," make it a hook, not a skill.

Contributing

Early days - issues and PRs welcome once it stabilizes. Each piece should be self-contained, documented, and (where it has a verifiable output) shipped with a tested script.

License

MIT - see LICENSE.