Part 1: What Are Skills?

Think of skills as portable instruction manuals for AI assistants. Instead of typing the same complex prompts over and over, you package them into reusable folders that any compatible AI agent can load on demand. Here's a mental model that helps: forget the word "Code" in Claude Code. Think of it as Claude Local or Claude Agent—a locally-running AI that performs tasks directly on your computer. Skills are how you teach this agent specialized expertise without bloating every conversation with instructions. A skill is simply a directory containing a SKILL.md file with metadata and instructions, plus optional scripts and resources. The AI reads only the skill's name and description at startup. When your request matches a skill's domain, the full instructions load into context. This "progressive disclosure" pattern keeps the AI's context window clean while enabling a massive library of capabilities.

What are AI agent skills and how do they work?

Skills are portable instruction packages that AI assistants load on demand. A skill is a folder containing a SKILL.md file with YAML frontmatter (name, description) and Markdown instructions. The AI reads skill metadata at startup, then loads full skill content only when your request matches the skill's domain.

How do I create my first SKILL.md file?

Create a folder in ~/.claude/skills/ with your skill name. Add a SKILL.md file with this structure: YAML frontmatter with name and description, followed by Markdown instructions. The description should include specific phrases that trigger the skill. Keep the file under 2,000 words and use imperative form ("Configure the server" not "You should configure").

What's the difference between skills in Claude Code vs OpenAI Codex?

Both use the same SKILL.md format through the agentskills.io open standard. Claude Code stores skills in .claude/skills/, while Codex uses .codex/skills/. Both implement progressive disclosure (metadata loads first, then full content when triggered). Skills are portable between both platforms with minimal modifications.

How many skills can I use at once without performance issues?

Keep active skill tokens under 20% of available context. For a 200K context window, that's 40K tokens for skills. In practice, 3-5 skills (5,000-15,000 tokens) works well for most tasks. Use budget-aware loading to preview skills before fully loading them.

Where can I find pre-built skills to download?

Three main sources: ABclaude-collection (300+ skills across 17 categories), claude-scientific-skills (125+ scientific skills), and claude-code-workflows (107 skills with multi-agent orchestration).

What's the difference between skills and plugins?

Skills are instruction packages (SKILL.md files with knowledge and procedures). Plugins are bundles that can contain skills, agents, commands, and tools. A plugin marketplace hosts multiple plugins. You install plugins, which give you access to their bundled skills.

Which plugin marketplace should I use for scientific research?

claude-scientific-skills by K-Dense. It has 125+ specialized skills covering bioinformatics (BioPython, Scanpy), cheminformatics (RDKit), clinical research, ML/AI, and drug discovery. Includes integrations with 26+ databases: PubMed, UniProt, NCBI, ChEMBL, DrugBank, ClinicalTrials.gov.

The Ultimate Guide to AI Skills: Claude Code, Desktop & GPT (2026)

AI DECK

Explore Tools

Curated By Experts

Free Forever

Find in 30 Seconds

Updated Daily

The Ultimate Guide to AI Skills: Claude Code, Desktop & GPT (2026) | AI Deck