it-department-skills

CI skills.sh Skills

Battle-tested global rules, portable skills, and a token-efficient skill-discovery system for AI coding agents (Claude Code, Cursor, Codex, Copilot — anything with a shell). Extracted from a production IT/agent stack and sanitized for public use.

Why "ultra performance"? Three levers:

  1. Strict global rules — one cross-tool ruleset (rules/AGENT_RULES.md) with a hard done-checklist, security defaults, and research protocol. Agents stop guessing and start verifying.
  2. Search-first skill discovery — once you have hundreds of skills, letting an agent ls skill directories burns thousands of tokens per session. skillfind gives ranked top-5 keyword search over an auto-built index; an optional PreToolUse hook makes search-first mandatory by denying raw enumeration.
  3. Compressed output modes — Absolute Mode (no filler, no hedging) ships in the rules; caveman (third-party, optional) compresses agent chat output ~75%.

Quick start

git clone https://github.com/vahidkaargar/it-department-skills.git
cd it-department-skills
python3 install.py --dry-run   # see what it would do
python3 install.py             # copies into ~/.ai, ~/.claude, ~/.agents (backs up existing)
python3 personalize.py         # fill in your machine/stack, toggle output modes

Installer and personalizer are pure-stdlib Python 3 — no bash/perl/shellcheck required, so this works identically on macOS, Linux, and Windows (use python instead of python3 on Windows if that's how it's aliased on your machine).

Or install individual skills via the skills.sh ecosystem CLI:

npx skills add vahidkaargar/it-department-skills --skill skills-discovery

Then in any agent session:

skillfind "redact secrets from log"   # -> pii-scrub SKILL.md path
rulesfind "output mode"               # -> absolute-mode spec

What's inside

Path What
rules/AGENT_RULES.md Cross-tool agent ruleset — priority order, standards, workflow, done-checklist
rules/absolute-mode.md Terse output grammar spec (eliminate/preserve lists, precedence)
skills/skills-discovery/ The mandatory search-first discovery skill (canonical doc)
skills/write-a-skill/ Author new skills correctly (progressive disclosure, <100-line SKILL.md)
skills/skill-tester/ Validate + score skill quality
skills/skill-security-auditor/ Audit third-party skills before installing them
skills/pii-scrub/ Redact secrets/PII from files before sharing
skills/remember/ Save durable knowledge to agent memory deliberately
skills/tech-debt-tracker/ Scan, score, and plan tech-debt remediation
skills/blindspot-check/ Red-team a big decision against 95 cognitive biases before committing
skills/thiel-style-converter/ Rewrite arguments in measured Zero-to-One strategic style (anti-fabrication built in)
skills/positioning-with-ekram/ Product/category positioning operator — diagnose before copy, eval-tested
skills/jscpd/ Copy-paste detector reference — run jscpd, read its AI-reporter clone output
discovery/skills-catalog/ skillfind/rulesfind CLI, index builder, optional localhost HTTP API
hooks/skills-discovery-guard.py PreToolUse hook: denies ls/find on skill dirs (opt-in)
templates/.ai/ Starter .ai/context.md + examples layout for the rules' context protocol
install.py / personalize.py Idempotent installer (copy + backup) and interactive personalizer — pure stdlib, cross-platform

This repo also dogfoods its own tooling for contributors (not copied to client machines by install.py):

Path What
CLAUDE.md / .rtk/filters.toml rtk (Rust Token Killer) instructions + project filters, via rtk init
.jscpd.json jscpd duplication-check config, enforced in CI (.github/workflows/ci.yml)

Design principles

  • One source of truth. Rules live in ~/.ai/rules.md; every tool (Claude, Cursor, Copilot) points at it. Edit once.
  • Search, never enumerate. Agents read ONE SKILL.md per task, found via keyword search over a prebuilt index — not directory listings.
  • Standalone copy installs. No symlinks — the installer copies everything into ~/.ai, ~/.claude, ~/.agents, so it works identically on any fresh machine and the clone can be deleted afterwards. Update by git pull + re-running python3 install.py (unchanged files are skipped, local edits backed up).
  • Nothing edits your settings silently. The guard hook is opt-in and the installer prints the settings.json snippet instead of merging it.

Uninstall

Plain copies — remove them and restore from the printed backup dir: ~/.it-department-skills-backup/<timestamp>/. Details: docs/uninstall.md.

Recommended companion plugins

Not part of this repo, not vendored — third-party Claude Code plugins that pair well with it. Install commands verified against Claude Code's /plugin docs; the @marketplace-name suffix is whatever that marketplace's own manifest declares, shown in the marketplace add command's own output if it ever differs from the table below.

Plugin What it adds Install Source
superpowers Brainstorming → TDD → systematic-debugging → writing-plans workflow discipline /plugin marketplace add anthropics/claude-plugins-official/plugin install superpowers@claude-plugins-official anthropics/claude-plugins-official
ashlr Token-efficient tool wrappers (read/edit/grep/bash) + cost tracking + multi-agent orchestration /plugin marketplace add ashlrai/ashlr-plugin/plugin install ashlr@ashlr-marketplace ashlrai/ashlr-plugin
ecc 100+ language/framework reviewers, build-fixers, and workflow commands (React, Go, Rust, Django, Kotlin, etc.) /plugin marketplace add affaan-m/ECC/plugin install ecc@ecc affaan-m/ECC
claude-mem Persistent cross-session memory — search past work, auto-capture, resume context /plugin marketplace add thedotmack/claude-mem/plugin install claude-mem@thedotmack thedotmack/claude-mem

Credits

  • caveman by Julius Brussee (MIT) — referenced as an optional companion, not vendored.
  • skills/write-a-skill builds on skill-authoring doctrine by Matt Pocock (MIT).
  • skills/blindspot-check, skills/thiel-style-converter, and skills/positioning-with-ekram by Soheil Momeni (MIT) — vendored with upstream LICENSE files; security-audited before inclusion.
  • skills/jscpd usage doc adapted from the jscpd project's own CLI reference (MIT).

License

MIT — see LICENSE.