Snapshot caveat: The skill ecosystems are new and evolving. The core idea is stable, but product names, plans, and which tools support the format all shift quickly, so re-verify specifics on each vendor's page. Reflects June 2026.
Agent Skills
Making AI Useful
8 min read
In Short
An agent skill is a folder you hand an AI agent that holds plain-language instructions plus optional extras like scripts, templates, or reference notes. The agent keeps only a one-line summary of each skill in mind and opens the full folder only when your task matches, an idea called progressive disclosure. Some skills are pure instructions, others bundle real code the agent runs for reliability. The point is to package an expert workflow once and reuse it, instead of re-explaining it every time.
01. What It Is
An agent skill is an organized folder of instructions and resources that an AI agent can discover and load when a task calls for it. At its simplest a skill is a directory containing one file named SKILL, which opens with a short name and a description, then plain-language instructions.
Anthropic, which introduced the format, compares building a skill to "putting together an onboarding guide for a new hire." Instead of a separate agent per task, you hand a general-purpose agent a written guide and it becomes a specialist. A skill is reusable, not a prompt you paste into one chat.
Anthropic's PDF skill makes this concrete. The folder holds a SKILL, two extra instruction files, and a small script that reads a PDF and extracts its form fields. The agent opens the form-filling instructions only when a task involves a form, and runs the script without loading its code or the PDF into memory.
02. Why It Matters
Language models do the same kinds of work over and over, and without skills you re-explain your preferences each time. A skill turns "explain it every time" into "package it once and reuse it." That is the payoff for a non-coder and a team alike.
A task you keep describing, like a monthly report's structure or a brand voice, becomes a skill the agent applies the same way every time, encoded once instead of living in someone's head. Because a project skill sits in a shared repository, a whole team's agents pick up the same guidance.
The build mechanics are light and covered separately (see making your own skill).
03. How It Works
Instructions plus optional bundled files
The one required piece of a skill is the SKILL file. It begins with two required fields, a name and a description, and the rest is plain language. Around it a skill can bundle more, including instruction pages, reference documents, templates, and runnable scripts. The open standard names common subfolders for these (scripts, references, assets), though none are required, and the smallest valid skill is a single SKILL.
Progressive disclosure, only name and description always loaded
The core idea is progressive disclosure, where the agent loads detail in stages. Anthropic describes it like a manual with a table of contents, then chapters, then an appendix.
At the first level the agent holds only the name and description of every installed skill, about 100 tokens each, loaded when it starts up. At the second level, when your task matches a description, it opens that skill's full SKILL, usually kept under about 5,000 tokens. At the third level and beyond, bundled files open only when needed. Because unused files never enter memory, Anthropic notes the material you can bundle is "effectively unbounded."
So you can install many skills without slowing the agent, since it only reads the one that fits.
The key distinction, instruction-only skills versus skills that bundle code
The distinction most people want is whether a skill carries code. The minimum skill is pure natural-language guidance, like a page on how your team structures a client report. Many skills never need code.
Other skills bundle a script the agent runs. Anthropic's guidance is to add code "for tasks where traditional programming is more reliable than token generation," work like sorting a list or parsing a PDF. When a skill includes a script, the agent runs it and only the output returns to context, never the source, so the result is deterministic and repeatable rather than improvised. This is also why a skill carries a real security cost, since an installed skill can run code on your behalf.
The ecosystems as of mid-2026
Skills started at Anthropic, but the instinct shows up under several names. Anthropic Agent Skills are the originator, launched on October 16, 2025 across Claude's apps at claude.ai, Claude Code, the Claude API, and the Claude Agent SDK at once. In December 2025 Anthropic published the SKILL format as an open standard, and it ships ready-made skills for office files like PowerPoint, Excel, Word, and PDF.
In Claude Code a skill lives in one of three places, a personal folder for every project, a project folder shared through version control, or an installed plugin. The agent loads the right one automatically, and you can also trigger one by typing a slash and its name. A skill can be left for the agent to invoke on its own, made manual-only for side-effect actions like a deploy, or kept as background knowledge.
For coding agents the parallel convention is AGENTS, "a README for agents." It is a plain Markdown file in a repository giving an agent build steps, tests, and conventions, with no required fields. It came from a cross-vendor effort, now stewarded by the Agentic AI Foundation under the Linux Foundation. OpenAI's Codex reads it before starting work, with global, repo-root, and per-folder files stacking, closest one winning. Cursor keeps its rules in a .cursor/rules/ folder and also reads AGENTS.
Different names, one instinct, all giving the agent reusable, on-demand instructions instead of repeating yourself.
04. Skills vs Tools vs MCP
These three sit at different layers.
A tool is one action the agent can take, like sending an email or running a database query (see tool use and function calling).
MCP is a standard way to connect the agent to outside data and tools in the first place (see MCP). A skill is the packaged know-how for doing a job, including which tools to use and when. Anthropic positions skills and MCP as complements, where MCP makes the connection and the skill supplies the workflow.
A system prompt is always on and applies to everything the agent does (see system prompts), while a skill stays out of the way until a task matches its description.
A subagent is a separate worker with its own memory and tools (see agents and agentic workflows), while a skill is not a worker but a capability the main agent or a subagent picks up.
05. Key Terms
| Term | Plain meaning |
|---|---|
| Agent skill | A folder you give an AI agent with plain-language instructions plus optional bundled files (scripts, templates, reference docs) that it loads only when a task matches. |
SKILL |
The one required file in a skill. It opens with a short name and description, then plain-language instructions. This is the open-standard file format Anthropic released. |
| Progressive disclosure | The agent always sees only a one-line summary of each skill and opens the full instructions and files only when the task calls for them. Keeps it fast and uncluttered. |
| Instruction-only vs code-bearing skill | An instruction-only skill is guidance in words. A code-bearing skill also bundles a script the agent runs, for jobs where a program is more reliable than writing the steps out. |
| Model-invoked vs user-invoked | Whether the agent decides to use a skill on its own, from the description, or you trigger it by hand, for example by typing a slash and its name. Many skills can be both. |
AGENTS |
A plain Markdown "README for agents" used by coding tools like Codex and Cursor to give an agent a project's build steps, tests, and conventions. A sibling convention to skills. |
| Subagent | A separate AI worker with its own memory and tools that the main agent hands a sub-task to. Different from a skill, which is a capability, not a worker. |
06. Common Misconceptions
"A skill is just a saved prompt."
A saved prompt is text you paste into one conversation. A skill is a reusable package the agent finds on its own and opens only when relevant, and it can carry files and runnable code.
"Installing lots of skills will slow the agent down or clog its memory."
Progressive disclosure keeps only a one-line summary of each skill, around 100 tokens, in working memory. The full instructions and files stay on disk until a task needs them, so reference material costs almost nothing until used.
"Skills are a Claude-only, Anthropic-only thing."
They began at Anthropic in October 2025, but the SKILL format became an open standard in December 2025 and now runs across several tools. The same instinct appears as AGENTS for coding agents and rules files in Cursor.
"Every skill contains code."
Many skills are pure natural-language instructions, like a house style guide. Code is optional, added only when a deterministic program beats having the agent generate the steps itself.
"A skill is the same as a tool, or the same as MCP."
They sit at different layers. A tool is one action. MCP connects the agent to outside tools and data. A skill is the know-how for doing the job, and Anthropic frames skills and MCP as complements.
"Skills are safe to grab from anywhere."
A skill can include scripts the agent will run, so installing one is closer to installing software than copying a prompt. Anthropic warns a malicious skill could exfiltrate data or take unintended actions, and advises sticking to trusted sources.