Agent Skills — The Cross-Vendor Format for LLM Agent Task Instructions
As of 2026-10-20, Agent Skills is the open standard Anthropic published in Dec 2025: SKILL.md needs only name and description, content loads in three levels, and Claude Code, claude.ai, the Claude API and OpenAI Codex / ChatGPT all read the same file.
What Skills Are: How They Differ from Plugins and MCP
A Skill is a directory containing SKILL.md (name and description in YAML frontmatter, task instructions in markdown) plus optional scripts, references, and assets. Anthropic calls them reusable, filesystem-based resources: the model matches your request against the description, then reads the body and the scripts it needs, and only a script's output enters the context. Different from Plugins (client-feature packages) and MCP (external tool protocol) — Skills aren't tools or client capabilities; they're 'how to do it' instruction bundles. One plugin can carry several Skills, and a Skill can reference MCP tools.
SKILL.md Spec: Required Fields and a Minimal Example
SKILL.md uses YAML frontmatter to declare name and description, with the body in markdown for instructions. In the spec only name and description are required; license, compatibility, metadata and allowed-tools are optional, and allowed-tools is still experimental. The official minimal example sets name to pdf-processing and describes extracting text and tables from PDF files, filling forms and merging documents, to be used when the user mentions PDFs, forms or document extraction. name is at most 64 characters, lowercase letters, numbers and hyphens only, with no leading or trailing hyphen; description is non-empty and at most 1024 characters. The directory is skill-name/SKILL.md plus optional scripts/, references/ and assets/. The body holds the steps for the model, for example: 1. Extract the files to process from the user message; 2. Call ./scripts/extract.py to output the result. The model loads this skill automatically when it sees a PDF-related request.
The Official Skills Repo and Partner Examples
The official anthropics/skills repository publishes a set of reference implementations, and its README sorts them into creative design, technical development, enterprise communication and document skills; docx, pdf, pptx and xlsx are the source-available versions behind Claude's document capabilities, while most of the rest are open source under Apache 2.0. On the platform side there are pre-built document skills: PowerPoint (pptx), Excel (xlsx), Word (docx) and PDF (pdf), ready to use on claude.ai and the API. The repo also ships a skill template and a copy of the spec, and in Claude Code you can register the repository as a plugin marketplace and install from it; the README's Partner Skills section currently highlights Notion's Claude skill. The community aggregator SkillsMP (skillsmp.com) describes itself as an independent community project that aggregates public GitHub sources to make skill discovery easier.
Cross-Vendor Compatibility and Enterprise Distribution
After Anthropic open-standardized Skills in Dec 2025, OpenAI's own docs state that Codex and ChatGPT read the same SKILL.md: Codex loads local skills from $CWD/.agents/skills, $HOME/.agents/skills and /etc/codex/skills, invoked explicitly with /skills or $, and can also trigger a skill by matching its description. The same SKILL.md runs in Claude Code, Codex CLI and ChatGPT — write once, deploy across vendors, though the surfaces do not sync: skills uploaded to claude.ai must be uploaded to the API separately, skills uploaded through the API are not available on claude.ai, and Claude Code skills are filesystem-based and separate from both. On the enterprise side, custom skills on claude.ai are individual to each user and Anthropic states that claude.ai does not support centralized admin management or org-wide distribution of custom Skills; Claude Enterprise organizations can turn on Skill content scanning for skills uploaded in claude.ai and Claude Cowork, while skills uploaded through the Skills API are not scanned, and an API request supports at most 20 Skills.
FAQ
Are Skills the same as Custom Subagents?
No. A custom subagent is a Claude Code dispatch unit with its own prompt, tools, and dispatch rules. A Skill is a higher-level instruction package that can be loaded by a main agent or any subagent. A subagent can reference multiple skills: the official way is the skills field, which preloads the full skill content into the subagent's context at startup, and the reverse also works — a skill with context: fork runs inside a forked subagent.
Can I use a Skill to enforce that the model avoids a behavior?
Not as a hard constraint — Skills are 'recommendations and instructions' that the model can still deviate from under pressure. Anthropic defines a skill as instructions and scripts, so whether Claude follows them is an evaluation question (the official checklist covers instruction following and triggering accuracy). For real restrictions, use the Claude Code mechanisms: disallowed-tools removes tools from the available pool while the skill is active, allowed-tools pre-approves tools, and disable-model-invocation: true stops Claude from invoking the skill on its own. Use Skills for 'how to do it right'; put hard rules elsewhere.
Can Codex and ChatGPT use the same SKILL.md format?
Yes. Skills are markdown text repositories — git clone and use locally: Claude Code reads them from ~/.claude/skills/ or .claude/skills/, Codex from .agents/skills directories. Paired with QCode's gateway to Claude/Codex/GPT models, the same SKILL.md works across vendors. OpenAI's docs say Skills build on the open agent skills standard: the SKILL.md file must include name and description, and Codex and ChatGPT start with each skill's name and description, then load the full SKILL.md when they decide to use it.
Use Agent Skills Across Vendors via QCode
QCode bundles Claude / GPT / Chinese model access — all compatible with the Agent Skills standard. Write once, run everywhere.
Start Your QCode PlanCreate a QCode account
One key across Claude, GPT and Chinese models, billed per token — live rates on /models.