Skip to main content

Skills

Skills extend what AI coding agents can do -- structured markdown files that define how an agent behaves in specific domains. Unlike opaque tool integrations, skills are readable, auditable, and customizable.

Three Ways to Use Any Skill

Every skill supports three invocation methods:

  1. Natural language -- just describe what you need (e.g., "design the system architecture")
  2. Slash command -- sw:skill-name in Claude Code (e.g., sw:architect)
  3. CLI keyword -- skill-name in Cursor, Copilot, and other AI tools (e.g., architect)

Start with Skills, Plugins & Marketplaces Explained for a deeper guide.

SpecWeave adds two complementary layers to the skill ecosystem, addressing the two fundamental questions every skill user faces:

  1. How do I make skills work for MY project? — Extensible Skills
  2. How do I know a skill is safe to install? — Verified Skills

The Two Layers​

Extensible Skills​

The customization layer. SpecWeave builds on Claude Code's native skill system with a clean separation:

  • Core instructions live in SKILL.md, stable and versioned
  • Your project-specific rules live in skill-memories/*.md

Skills self-load their customizations using dynamic context injection (Claude Code's built-in !`command` syntax) — a shell one-liner that reads your preferences before the skill executes. Skill memories are loaded from project, personal, and global directories, with project-level overrides taking priority over global defaults.

The result: you correct Claude once ("use React Hook Form, not useState for forms"), and that preference is applied automatically in every future session. No reminders needed.

Read more about Extensible Skills

Verified Skills Standard (V-Skills)​

The trust layer. A graduated security certification system that filters dangerous skills before they reach your codebase.

Snyk's ToxicSkills study (February 2026) found that 36.82% of 3,984 publicly available skills contained security flaws, including 76 confirmed malicious payloads — credential theft, crypto miners, and prompt injection designed to persist across sessions. No existing platform had meaningful security scanning.

The Verified Skills Standard introduces three-tier certification:

TierMethodCostSpeed
Scanned52 pattern checks + structural validationFree< 500ms
VerifiedTier 1 + LLM intent analysis~$0.03/skill5-15s
CertifiedTiers 1+2 + human security review + sandbox$50-200/skill1-5 days

The registry at verified-skill.com provides a trusted source for browsing and submitting skills, with the vskill CLI for command-line access (npx vskill, bunx vskill, pnpx vskill, or yarn dlx vskill).

Read the full Verified Skills Standard

How They Complement Each Other​

                BEFORE INSTALL                    AFTER INSTALL
          ┌─────────────────────┐          ┌─────────────────────┐
          │  Verified Skills    │          │  Extensible Skills  │
          │                     │          │                     │
          │                     │          │                     │
          │  "Is this skill     │    ──►   │  "How do I make     │
          │   safe to use?"     │          │   this skill work   │
          │                     │          │   for MY project?"  │
          │  3-tier trust       │          │  Injection +        │
          │                     │          │  memories           │
          └─────────────────────┘          └─────────────────────┘

A skill can be both verified (passed security certification) and extensible (customizable via skill memories). These layers address different stages of the skill lifecycle:

  • Verified Skills answers: Should I trust this skill? — evaluated before installation
  • Extensible Skills answers: How do I adapt this skill? — customized after installation

Explore Further​