66+ focused skills. Skills come in three formats: (1) core skills using the standard template — Identity → Stack Defaults → Decision Framework → Anti-Patterns → Quality Gates (~75 lines); (2) Android/platform reference skills — full implementation guides with working code patterns; (3) third-party vendor skills with their own YAML schema (e.g. Google, Azure). Dense enough to be useful, lean enough not to bloat your context window.
Skills load into your AI's context window on every invocation. Bloated skills waste tokens and dilute focus. This library is optimized for signal density:
| Before (v1) | After (v2) | |
|---|---|---|
| Files | 47 (35 skills + 12 "super agents") | 65+ skills |
| Total lines | 19,207 | 2,531 |
| Est. tokens (full load) | ~240,000 | ~37,000 |
| Avg lines/skill | ~400 | 72 |
| Tokens per invocation | ~3,700 | ~1,050 |
−87% lines. −84% tokens per invocation. Same decision quality — the template forces every skill to include if/then decision rules, anti-patterns, and quality gates rather than prose explanations.
Every skill answers three questions: What tool/approach? (Stack Defaults), When and why? (Decision Framework), What to avoid? (Anti-Patterns).
# Full collection
git clone https://github.com/drewid74/ai_skills.git ~/.claude/skills/ai_skills
# Or cherry-pick individual skills
cp -r ai_skills/docker-selfhost ~/.claude/skills/Works with Claude Code, Cursor, Codex CLI, Windsurf, and any tool that reads SKILL.md files.
| Skill | Lines | What It Does |
|---|---|---|
| docker-selfhost | 68 | Docker Compose stacks, self-hosted services, volumes, reverse proxy, TrueNAS SCALE |
| truenas-ops | 76 | ZFS pool management, TrueNAS datasets, replication, snapshots, permissions |
| proxmox-k3s-infra | 78 | Proxmox VMs/LXC, cloud-init templates, GPU passthrough, K3s clusters, GitOps |
| deploy-pipeline | 82 | rsync/SSH deploys, health checks, rollback, secrets hygiene, ZFS snapshots |
| infrastructure-as-code | 78 | Terraform/OpenTofu, Ansible, state management, drift detection, secrets handling |
| llm-inference-stack | 79 | Ollama, vLLM, LiteLLM routing, VRAM sizing, quantization formats |
| service-integration | 62 | Webhooks, n8n, message queues, notification pipelines, Traefik |
| orchestration-scaffold | 555 | OpenCode + LangGraph + Ollama + MCP scaffold: harness setup, agent wiring, SLIs, VS Code integration, TrueNAS/Docker Compose |
| Skill | Lines | What It Does |
|---|---|---|
| full-sdlc | 66 | Project scaffolding, branching, testing strategy, CI/CD, release management |
| code-reviewer | 62 | Correctness, security, edge cases, performance — Python/JS/Go/Bash |
| cicd-pipeline | 84 | GitHub/Forgejo Actions, runners, caching, container registries, semantic release |
| github-workflow | 63 | Repo scaffolding, branch protection, PR workflow, issue management |
| quality-test-engineer | 79 | pytest/Jest/Playwright/k6, flaky test diagnosis, CI integration, coverage |
| database-architecture | 77 | Schema design, indexing, EXPLAIN ANALYZE, ORMs, migrations, connection pooling |
| api-integration | 77 | REST/GraphQL, OAuth/JWT, retry/circuit breaker, rate limits, webhooks |
| Skill | Lines | What It Does |
|---|---|---|
| ai-systems-architect | 77 | RAG vs fine-tuning, agent design, MCP, context budgeting, eval pipelines |
| mcp-server-dev | 68 | FastMCP/TypeScript MCP servers, tool design, transports, Docker packaging |
| training-pipeline | 75 | LoRA/QLoRA fine-tuning, hyperparameters, DeepSpeed, experiment tracking |
| federated-memory | 68 | Agent persistent memory, vector stores, Qdrant/pgvector, cross-session recall |
| ai-skills-dev | 67 | SKILL.md authoring, trigger optimization, template iteration |
| Skill | Lines | What It Does |
|---|---|---|
| security-engineer | 80 | OWASP Top 10, secrets scanning, container hardening, SAST, supply chain, CVE triage |
| sre-operations-lead | 79 | Prometheus/Grafana/Loki, SLOs, alert design, incident response, runbooks |
| Skill | Lines | What It Does |
|---|---|---|
| data-engineering | 71 | ETL/ELT pipelines, pandas/polars/DuckDB, idempotency, schema evolution |
| sigint-osint-feeds | 78 | APRS, ADS-B, GDELT, OSINT aggregation, PostGIS, worker pipelines |
| archivebox-knowledge | 72 | Web archival, Paperless-NGX, content extraction, knowledge base pipelines |
| research-analyst | 67 | Source-grounded research, triangulation, confidence scoring, threat intelligence |
| Skill | Lines | What It Does |
|---|---|---|
| web-performance-a11y | 80 | Core Web Vitals, WCAG 2.1/2.2 AA, Lighthouse, asset optimization |
| frontend-design-pro | 64 | Design systems, Tailwind v4, Shadcn/Radix, accessibility audits, mobile-first |
| browser-automation | 73 | Playwright, scraping, anti-bot handling, page monitoring, e2e testing |
| Skill | Lines | What It Does |
|---|---|---|
| content-strategy | 66 | Technical writing, READMEs, API docs, changelogs, SEO |
| productivity-automation | 80 | Cron, batch jobs, file processing, backup automation, monitoring |
| gws-assistant-pro | 63 | Gmail, Calendar, Sheets, Drive automation |
| Skill | Lines | What It Does |
|---|---|---|
| project-orchestrator-pro | 64 | Multi-step project coordination, task decomposition, agent delegation |
| sequential-thinking-pro | 62 | Root cause analysis, architecture review, structured decision-making |
| repo-auditor-pro | 67 | Repository health, docs quality, broken links, release readiness |
| ham-radio-network | 79 | Antenna math, CHIRP/DMR, FT8/APRS, AREDN mesh, VLAN/firewall design |
Full implementation reference skills with working Kotlin/Compose code patterns. Each covers the idiomatic approach, common pitfalls, and API-level caveats.
| Skill | What It Does |
|---|---|
| android-auto | Android Auto / Automotive OS: Car App Library templates, CarAppService/Session lifecycle, voice interaction, distraction optimization rules, navigation and POI templates, AAOS native targeting |
| android-ble-hardware | BLE hardware integration via Nordic BLE library: UART/GATT service discovery, callbackFlow scanning, multi-device type detection, automatic reconnect, Android 12+ permission model, binary packet parsing |
| android-compose-realtime | High-frequency real-time UI in Jetpack Compose: recomposition avoidance with granular state and derivedStateOf, Canvas-based gauges and sliding-window graphs, StateFlow → Compose at 25Hz+, performance profiling |
| android-foreground-service | Long-running foreground services for sensor/BLE/GPS/audio collection: Android 14+ foregroundServiceType enforcement, START_STICKY null-intent safety, intent-based IPC, source-priority fallback, clean shutdown |
| android-on-device-ml | On-device ML: ML Kit (text recognition, object detection, pose), TensorFlow Lite custom models with GPU/NNAPI delegates, MediaPipe Tasks API, CameraX + ML pipeline, model asset delivery |
| android-security | Android security fundamentals: EncryptedSharedPreferences/EncryptedFile, KeyStore-backed key generation, BiometricPrompt + Credential Manager passkeys, OkHttp certificate pinning, Play Integrity API, R8 rules |
| android-telemetry-pipeline | High-frequency sensor pipeline (10–100Hz): hybrid Room + CSV/binary storage, StateFlow streaming to Compose ViewModel, atomic session lifecycle, telemetry normalization across BLE/GPS/IMU sources, post-session analysis |
| android-video-capture | CameraX video and photo capture in Compose: lifecycle-aware camera setup, VideoCapture use case with MediaStore output, simultaneous preview + record + capture, camera selector, torch control, recording state management |
Every skill follows the same structure (~75 lines):
---
name: skill-name
description: "Use this when: [action/problem phrases ≤20, intent-first]"
---
## Identity
You [role]. [Core principle]. Never [hard constraint].
## Stack Defaults
| Layer | Choice | Why |
## Decision Framework
### When to X
- If [condition] → [action]
## Anti-Patterns
| Don't | Why | Do Instead |
## Quality Gates
- [ ] GateDescriptions use action/problem phrases ("my pipeline is failing", "set up CI") rather than tool names — this is how the routing system finds the right skill for natural-language prompts.
Decision density over prose. Every skill has explicit if/then decision rules and an anti-patterns table. The LLM gets when to use X vs Y — not just what X is.
Context-efficient. At ~1,050 tokens per skill invocation (down from ~3,700), these load fast and leave room for your actual code in the context window.
Universal, not personal. Placeholders (<NAS_IP>, <API_KEY>, <POOL_NAME>) throughout. No hardcoded infrastructure — patterns that work across any setup.
skill-name/
├── SKILL.md # Required: YAML frontmatter + ~75 lines of structured instructions
└── scripts/ # Optional: helper scripts referenced in SKILL.md
The YAML frontmatter controls when the skill auto-invokes:
---
name: docker-selfhost
description: "Use this when: my container keeps restarting, set up a self-hosted service,
my volumes are not persisting, deploy with Docker Compose, my container won't start"
---Descriptions use intent phrases ("my container keeps restarting") not tool names ("Docker, Compose, volumes"). The routing is semantic — natural-language problem descriptions match better than keyword lists.
| Catalog | Description |
|---|---|
| claude_capabilities_catalog.md | Tools, MCPs, skills, and triggers in Claude |
| chatgpt_capabilities_catalog.md | ChatGPT equivalent capabilities |
| gemini_capabilities_catalog.md | Gemini capabilities |
| cross_agent_skills.md | Audit prompt to generate a comparable catalog for any platform |
git clone https://github.com/drewid74/ai_skills.git ~/.claude/skills/ai_skillscp -r ai_skills/docker-selfhost ~/.claude/skills/
cp -r ai_skills/security-engineer ~/.claude/skills//skills list # shows all loaded skills
/skills reload # pick up new skills added mid-session
/docker-selfhost # invoke a skill explicitly by name
Skill directory paths vary by tool. For Claude Code:
~/.claude/skills/. For Copilot CLI:~/.copilot/skills/. Check your tool's documentation.
Skills are plain markdown with a fixed template. The bar is one skill, ~75 lines:
- Fork the repo
- Create
your-skill-name/SKILL.mdusing this structure:## Identity— role + core principle + hard constraint (3 lines)## Stack Defaults— 4–8 row table: Layer | Choice | Why## Decision Framework— 2–4 if/then sections with→bullets## Anti-Patterns— 4–6 row table: Don't | Why | Do Instead## Quality Gates— 4–6 checkboxes
- Write the description as action/problem phrases:
"Use this when: my X is broken, set up Y, how do I Z" - Open a PR with 2–3 example prompts that trigger your skill
Quality bar: Does it have real if/then decision rules? Does the anti-patterns table catch the top mistakes? Would a developer trust it at 2am during an incident? If yes, it belongs here.
MIT — use these skills however you want, commercially or otherwise.