Agentic Skills

An agentic skill is a modular capability package that extends what an AI coding agent can do. At its simplest, it's a SKILL.md file — a markdown document with YAML frontmatter (name, description, metadata) plus natural-language instructions — that an agent reads and follows when the right context arises.

The format was introduced by Anthropic when they open-sourced the Agent Skills specification on December 18, 2025. Within weeks, it was adopted by 26+ platforms including OpenAI Codex, GitHub Copilot, VS Code, Cursor, Gemini CLI, Windsurf, Manus, and Amp. Simon Willison called it “deliciously tiny” — and that minimalism is exactly why it spread so fast.

A SKILL.md uses progressive disclosure: metadata loads first (~100 tokens), then full instructions (under 5K tokens), then resources (scripts, references, assets) are pulled on demand. This prevents skills from overwhelming the agent's context window.

The .skill ZIP package bundles everything together: a SKILL.md file plus optional scripts/, references/, assets/, and templates/ directories. This is the distribution format used by skill marketplaces and team repositories.

Every platform evaluates skills internally, but none exports a portable trust score. There is no npm audit for agent skills. No Lighthouse score. No quality gate. The ecosystem has ~100,000 published skills and growing — with no immune system.

The ClawHavoc security incident (February 2026) exposed 824 malicious skills on ClawHub alone, with Bitdefender confirming 17% of audited skills contained malware. These weren't edge cases — they were skills with GitHub stars, install counts, and apparent legitimacy.

Hugging Face's upskill research revealed something arguably worse: skills can degrade agent performance. Curated skills improved model output by +16.2 percentage points on average, but self-generated skills actually decreased performance by -1.3 points. A skill optimized for Claude Opus may actively harm Claude Haiku. Without per-model quality scoring, developers deploy capability-degrading instructions and have no way to know.

Claude's own documentation warns developers to “exercise extreme caution” with untrusted skills — but provides no mechanism to actually verify that trust. The gap between warning and tooling is where risk lives.

The Yawn Skill Trust Score evaluates every skill file across 35+ checks organized into four weighted dimensions. The score produces a composite 0–100 value displayed as an F-to-A+ letter grade — universally understood, screenshot-worthy, and competitive.

How scoring works

Each dimension score is the percentage of checks passed within that dimension. The composite score uses weighted aggregation: Functional × 0.25 + Quality × 0.20 + Safety × 0.30 + Maintenance × 0.25.

Safety floor: If the safety score falls below 30, the composite is capped at D regardless of how well other dimensions score. An unsafe skill that “works” is worse than a safe skill that doesn't yet.

The heuristic engine runs entirely in the browser — zero API calls, instant results, no file upload. The optional Deep Scan sends content to an LLM for semantic analysis of instruction clarity, cross-model compatibility, and hidden safety gaps that static checks cannot detect.

The missing piece in every existing agent skill standard is holonic composition — skills that contain sub-skills, forming a navigable tree. In the .yawn system, the skill holarchy is simultaneously:

• Composition — which skills are needed to compose a capability
• Navigation — the URL path maps directly to the tree position
• Permission — depth in the tree determines autonomy requirements

Each skill tracks a loop status: green (loop closed — active, handler, evidence), yellow (partial — exists but incomplete), red (gap — stub, planned, or blocked). Unmatched slots are lacunae — explicitly identified gaps that the system surfaces for resolution.

The EU AI Act's high-risk system requirements take full effect August 2, 2026 — demanding documented risk management, technical transparency, human oversight, and conformity assessments. Penalties reach €35M or 7% of global revenue. For enterprises deploying agents with skills in regulated industries, every consumed skill becomes part of the compliance surface.

Gartner predicts over 40% of agentic AI projects will be canceled by end of 2027 due to escalating costs, unclear business value, or inadequate risk controls. The “inadequate risk controls” signal is the market opening for trust infrastructure.

The AI agent market reached $5.4 billion in 2024 and is projected to hit $236 billion by 2034 (45.8% CAGR). AI agents could mediate $3–5 trillion in global consumer commerce by 2030 according to McKinsey. The ecosystem has ~100,000 published skills, growing exponentially, with no quality immune system.

The FICO analogy is precise: credit scoring transformed lending from guesswork to data-driven risk assessment. Trust scoring will do the same for agentic AI. The first trust layer to become the default — the way npm audit became default, the way Dependabot became default on GitHub — wins the market.

The window is approximately 12–18 months before the market consolidates. The message that makes non-adoption feel irresponsible is straightforward: “Your AI agents have access to your codebase, your APIs, and your customer data. They're consuming skills published by strangers with no proof of quality. Would you run npm packages without npm audit?”