AI Agent Skills — Helicopter View

The Problem

LLMs Know Facts, Not Procedures

LLMs have vast semantic knowledge but lack procedural knowledge — the step-by-step "how to do things" that professionals build through experience.

System prompts are static, per-session
RAG retrieves facts, not workflows
Fine-tuning is slow and not composable

What Is a Skill

A Markdown File That Teaches an Agent How to Do Things

A SKILL.md file in a well-known directory (.claude/skills/). YAML frontmatter carries metadata. The markdown body holds step-by-step instructions. Optional folders for scripts, references, and assets.

Portable Composable Human-Readable Version-Controlled

---
name: deploy-to-vercel
description: Deploy a Next.js
  app to Vercel with env setup
---

## Instructions
1. Check for vercel.json
2. Verify env vars are set
3. Run vercel --prod

Cognitive Parallel

The Missing Third Memory

Semantic Facts = RAG

Episodic Experience = History

Procedural How-to = Skills

A surgeon knows anatomy (semantic), remembers a hard case (episodic), and knows how to cut (procedural).

Progressive Disclosure

3-Tier Loading

Tier 1 — Metadata only
~100 tokens/skill at startup

Tier 2 — Full instructions
Loaded on match, <5K tokens

Tier 3 — Resources
Scripts and assets on-demand

50x context reduction

Knowledge Stack

Where Skills Fit

Skills — Procedures

MCP — Tool Access

RAG — Factual Knowledge

Fine-tuning — Model Weights

MCP = what tools exist. Skills = when and how to use them.

Platform Adoption

7 Platforms, 6 Months

Claude Code Late 2025

OpenAI Codex Feb 2026

Cursor Q1 2026

Windsurf Q1 2026

MS Agent Framework 2026

Gemini CLI 2026

OpenClaw 2026

Security & Trust

ClawHavoc: 1,184 Malicious Skills

Security researchers found 1,184 malicious skills on ClawHub. Six attack categories identified:

Prompt Injection Tool Poisoning Malware Credential Leak Untrusted Content Toxic Flows

Knowledge Stack Comparison

Four Layers of Agent Knowledge

Layer	Knowledge Type	Update Speed	Portability	Context Cost
Fine-tuning	Model weights	Weeks	None	Zero
RAG	Factual retrieval	Minutes	Medium	Medium
MCP	Tool access	Instant	High	Low
Skills	Procedures	Instant	Highest	~100 tokens

Skill Loading Pipeline

How an Agent Resolves a Skill at Runtime

Anti-Patterns

What Breaks Skills

Monolithic 10K+ token skills
Loading all content at startup
Blind trust of community skills
Hardcoding tool names
Never updating procedures

Adoption Timeline

From One Platform to Industry Standard in 6 Months

Claude Code> OpenAI Codex> Cursor> Windsurf> MS Agent Framework> Gemini CLI> OpenClaw

Late 2025 — Claude Code ships .claude/skills/ with SKILL.md format. First platform to formalize procedural memory for AI agents.

Q1 2026 — Four more platforms adopt compatible formats. Skills become a de facto standard, like Dockerfile for containers.

2026 — OpenClaw launches ClawHub, a community registry. Microsoft and Google integrate skills into their agent frameworks.