The world's most comprehensive open-source directory of AI work agent skills.
527 skills · 86 agent roles · 15 departments · production-grade depth
Every skill ships with executable workflows, scoring rubrics, output templates, anti-patterns, and cross-references — written at the depth of the best human practitioners, not generic AI advice.
Most AI agent prompts tell an agent what to do. Skill OS tells it how the best people actually do it.
Each skill is a structured markdown file containing:
- Workflow — imperative, numbered steps with named deliverables
- Anti-patterns — what to avoid and why, not just what
- References — named frameworks, decision tables, scoring rubrics (loaded on demand)
- Assets — production-ready output templates with realistic placeholders
- Related skills — bidirectional cross-references with rationale
Skills are scoped to an agent role (e.g. account-executive, sr-backend-developer, pmm-product-marketing-manager) so the context is always domain-specific, never generic.
| Department | Agents | Skills |
|---|---|---|
| Engineering | 12 | 100 |
| Product | 5 | 82 |
| Marketing | 12 | 60 |
| Data & Growth | 4 | 39 |
| Design | 8 | 43 |
| Finance | 5 | 32 |
| Legal & Compliance | 4 | 36 |
| Agent Operations | 6 | 27 |
| Sales | 7 | 27 |
| Customer Success | 6 | 21 |
| Customer Support | 2 | 9 |
| Technical Operations | 3 | 9 |
| Revenue Operations | 1 | 6 |
| Applied Research | 1 | 5 |
| Account Management | 2 | 5 |
| Implementation | 2 | 4 |
| Total | 80 | 528 |
Full org chart: AGENTS.md
skill-os/
├── agents/ # All skills live here
│ └── <agent-slug>/
│ └── <skill-slug>/
│ ├── SKILL.md # Skill definition (YAML frontmatter + 9 sections)
│ ├── references/ # Frameworks, rubrics, checklists
│ ├── assets/ # Output templates
│ ├── examples/ # Input/output pairs
│ └── scripts/ # Executable helpers
├── departments/ # Department ethos profiles
│ └── <dept>/
│ └── ideal-<dept>.md # What excellent looks like in this department
├── _shared/ # Cross-department resources
├── scripts/
│ └── validate.py # Validation script (zero external deps)
├── allowed-tools.yaml # 4-level tool access policy
├── ETHOS.md # Operating philosophy
└── status.md # Enrichment progress tracker
Each SKILL.md is self-contained. Load it into your agent's context and the agent will follow the workflow, apply the anti-patterns, and produce output matching the template.
Example — loading a skill into a Claude prompt:
You are an Account Executive. Execute the following skill:
<skill>
[contents of agents/sales/account-executive/meeting-prep-builder/SKILL.md]
</skill>
Context: [your deal context here]
Progressive loading — the skill body stays within strict word limits (500/1,000/1,500 by complexity). Reference files and templates load on demand when the agent needs them, so they never bloat the base prompt.
Context lookup chain: skill → agent → department → _shared/. Most specific wins.
Every skill has YAML frontmatter and 9 sections:
---
name: Meeting Prep Builder
description: Builds a battle-ready meeting brief so the AE walks in knowing the deal cold.
department: Sales
agent: account-executive
version: "1.0"
complexity: medium
related-skills:
- deal-qualifier
- sales-proposal-builder
---Sections: title · agent header · skill description · when to use · workflow · anti-patterns · output · related skills
Full spec: specs/001-enrich-skill-directory/contracts/skill-template.md
Connect Skill OS's 528 skills to Paperclip, the open-source AI agent orchestration platform.
Set instructionsFilePath in your Paperclip agent's adapterConfig to any SKILL.md path — the agent will follow that skill's workflow for every task it receives.
See paperclip/QUICKSTART.md for setup, routing tables, [GATE] handling, and progressive loading guidance.
# Validate structure and schema
python3 scripts/validate.py # Full repo
python3 scripts/validate.py agents/<agent>/ # One agent
python3 scripts/validate.py agents/<agent>/<skill>/SKILL.md # One skill
# Score quality (triggers, depth, references, assets, sections)
python3 scripts/score.py # Top 20 leaderboard
python3 scripts/score.py --bottom 20 # Worst 20 — contributor targets
python3 scripts/score.py --summary # Average score per department
python3 scripts/score.py --json # Machine-readable outputNo external dependencies. Requires Python 3.10+.
- Pick an agent from the org chart
- Check
status.mdfor coverage gaps - Follow the format in
CLAUDE.mdand the skill template - Validate before opening a PR
- One skill per commit:
Enrich skill: <skill-slug> for <Agent Role>
Read ETHOS.md before writing — it defines the depth standard.
A skill that does one thing with genuine domain expertise is worth more than ten skills that skim the surface.
Skills in this repo are written at practitioner depth — not what a generalist imagines a job looks like, but how the best people in that role actually work. Every framework is named. Every anti-pattern has a rationale. Every workflow step has a deliverable.
See ETHOS.md for the full philosophy.
MIT