Claude Skills: the open standard 32+ AI tools now support (2026 guide)

Agent Skills started as Anthropic's internal format for letting Claude load specialised instructions per task. Eight months later, it's the only AI-tool extensibility format adopted by all four major AI IDE vendors — Anthropic, OpenAI, Google, and Microsoft. This is what Skills are, why the open standard matters, and how to write one.

The 30-second mental model

A Skill is a folder. The folder has a SKILL.md file that tells the model what the Skill does and when to use it, plus optional scripts, templates, and reference docs. The model loads the Skill when a matching task arrives — no manual selection required.

my-skill/
├── SKILL.md          # Required: instructions + metadata
├── scripts/          # Optional: Python/JS/Bash executable code
│   └── helper.py
├── references/       # Optional: deeper docs the model can load
│   └── api-reference.md
└── assets/           # Optional: templates, schemas, examples
    └── template.docx

The SKILL.md frontmatter:

---
name: excel-financial-report
description: |
  Use this skill when the user asks to generate a financial report
  spreadsheet. Produces a multi-sheet .xlsx with summary, detail,
  and pivot tables.
---
 
# Instructions
 
When invoked, generate an Excel financial report by:
1. Reading the input data from `references/data-schema.md`
2. Using `scripts/generate_report.py` to assemble the .xlsx
3. Returning the file path to the user

That's the whole format. Read the description; if Claude thinks your prompt matches, the full Skill loads. Progressive disclosure — the metadata is always in context, the full body only when needed.

Why this caught on (the open standard moment)

October 16, 2025: Anthropic shipped Agent Skills for Claude.

December 18, 2025: Anthropic released the SKILL.md spec at agentskills.io↗ as an open standard, donating governance to the Agentic AI Foundation under the Linux Foundation.

Within 48 hours:

• Microsoft integrated SKILL.md into VS Code via Copilot
• OpenAI added support to Codex CLI

By March 2026, 32+ tools had adopted the format:

• Google Gemini CLI
• JetBrains Junie
• AWS Kiro
• Cursor
• GitHub Copilot
• Continue.dev
• Aider
• ...

This is the first time the AI IDE category has converged on a single extension format. Pre-Skills, each tool had its own:

• VS Code: extensions + Copilot custom instructions
• Cursor: .cursor/rules + custom prompts
• Claude Code: ~/.claude/agents/*.md subagents
• Cline: custom MCP servers

Post-Skills: write once in SKILL.md, the same Skill loads in any compatible tool.

What Skills are good for

Three high-value patterns:

1. Domain knowledge packaging

Your company has internal conventions, naming patterns, deployment processes that aren't in the model's training data. Package them as a Skill, commit to the repo, every developer using any compatible AI tool gets the same baseline.

---
name: acme-deployment
description: |
  Use this skill when the user asks to deploy or release a service
  in the Acme codebase. Includes our staging→prod gates, runbook
  links, and on-call expectations.
---

2. Output formatting + templates

If your team always produces certain artifact types (architecture docs, RFCs, runbooks, PR templates), encode the format once:

---
name: acme-rfc
description: |
  Use this skill when the user asks to write an RFC. Produces a
  .md file in our standard RFC template with sections: Problem,
  Non-goals, Proposal, Alternatives Considered, Rollout, Risks.
---

3. Tool wrappers

Common scripts that should always be invoked the same way — wrap them in a Skill so the model can call them consistently:

---
name: deploy-staging
description: |
  Use this skill when the user asks to deploy to staging. Invokes
  scripts/deploy_staging.sh after running pre-flight checks.
---

Where Skills live

Three deployment surfaces:

Anthropic ships four pre-built Skills out of the box: docx, pptx, xlsx, pdf — these are what powered Claude's document creation features starting September 2025.

Production results: 87.5% time reduction

The most-cited deployment number: Rakuten reported 87.5% workflow time reduction using Skills for financial workflows. Full-day tasks became one-hour tasks.

The mechanism: Skills compose. The model picks the right combination automatically. A "generate quarterly financial report" prompt might trigger:

1. xlsx skill (build the spreadsheet)
2. acme-financial-templates skill (apply company formatting)
3. pdf skill (export the executive summary)

Pre-Skills, the user would orchestrate these by prompting through each step. Post-Skills, one prompt fires the whole chain.

How to write one (5-minute version)

# 1. Create the skill folder
mkdir -p ~/.claude/skills/my-skill
 
# 2. Write SKILL.md
cat > ~/.claude/skills/my-skill/SKILL.md << 'EOF'
---
name: my-skill
description: |
  Use this skill when the user asks to do X. Produces Y.
---
 
# Instructions
 
When invoked:
1. Read the input
2. Do the thing
3. Return the result
EOF
 
# 3. (Optional) Add scripts
mkdir ~/.claude/skills/my-skill/scripts
cp my_script.py ~/.claude/skills/my-skill/scripts/
 
# 4. Reload Claude Code (or open a new session)
# Skills auto-discover from ~/.claude/skills/ on startup
 
# 5. Test it
claude "do X for me"
# Claude should invoke my-skill automatically

In Claude Code 2.1+ (January 2026), Skills hot-reload — no restart required after editing.

Composability + the description field

The single biggest factor in whether your Skill gets invoked correctly: how specific your description is.

Bad:

description: Helps with code review.

Good:

description: |
  Use this skill when the user asks to review a TypeScript pull
  request that touches the Acme payments codebase. Applies our
  PCI-DSS-compliance checklist and the Acme code review template.

The model uses the description to decide which Skill (or skills, plural — they compose) to load. Vague descriptions cause silent non-invocation. The harder you specify the when, the more reliable the Skill becomes.

Common gotchas

• Description too long. Claude.ai caps descriptions at 200 characters. The Agent Skills spec allows 1024. Test on the surface you care about.
• Skills competing. If two Skills have overlapping descriptions, Claude picks one — usually correctly, but inspect the chain of thought to verify.
• No telemetry. Skills don't surface invocation metrics out-of-the-box. Track usage via Claude Code's claude-skills.log file or via API request inspection.
• Skill version drift. When teams share Skills via git, version the SKILL.md frontmatter (version: 1.2.0) and document breaking changes.

Skills + Anvat

Anvat is an Anthropic-compatible API gateway. Skills work end-to-end through Anvat exactly as they do directly against api.anthropic.com:

• Claude Code via Anvat: Skills installed in ~/.claude/skills are loaded by Claude Code regardless of the upstream endpoint. Anvat is transparent to the Skills layer.
• API via Anvat: The /v1/skills endpoint is passed through. Custom Skills uploaded via the API work through Anvat with the same beta headers (code-execution-2025-08-25, skills-2025-10-02).

The cost benefit: Skills frequently produce long output (full spreadsheets, multi-page documents). Output tokens are the expensive side. Anvat's 30% off output rate compounds with Skills' productivity gains.

Bottom line

Skills are the AI-tool extensibility format that won the standardisation race. If you're building anything with Claude Code, Cursor, or any of the 32+ Skills-compatible tools, learning the format is now a 30-minute prerequisite. The ceiling isn't the model — it's how well your tooling captures the team's expertise.

Building coding agents in 2026 → Claude Code MCP setup guide →