On this pageJump to a section
Every operator I know who uses AI seriously has the same file. It started as a paragraph. Then a page. Then six pages of instructions about tone, formatting, offers, edge cases and "never do this again," pasted at the top of every conversation.
I call it the mega-prompt, and it has a short life. It gets long, it gets stale, nobody remembers why line 41 is there, and the AI reads all of it every time whether the task needs it or not. Anthropic's Agent Skills, introduced October 16, 2025, are a direct answer to that problem.1
What is an Agent Skill, in plain English?
Anthropic's own definition is short: skills are folders that include instructions, scripts and resources that Claude can load when needed.1
Think of a skill as one binder on a shelf. The binder has a label, a page of instructions on how to do one job, and maybe a template or a small script. When you ask for that job, the assistant takes down the right binder. When you ask for something else, the binder stays on the shelf.
Skills are folders that include instructions, scripts, and resources.
The announcement lists a few properties worth noting:1
- On-demand loading: Claude scans available skills and loads only what a task needs.
- Composable: skills can be combined, and Claude coordinates which ones to use.
- Portable: the same skill format works in Claude apps, Claude Code and the API.
- Can include code: for tasks where running a script is more reliable than generating text.
- You ask for a jobClaude scans the available skills
- One skill matchesOnly the minimal information and files load
- The steps runWith a script where code is more reliable
- The binder goes backNothing else clogs the conversation
Why does the mega-prompt break?
Three reasons, and each one gets worse as your business grows.
It is always on
A prompt you paste every time takes up room in the conversation whether or not it is relevant. Anthropic's context engineering guidance calls out context rot: recall accuracy tends to drop as the context window fills, so more context is not automatically better.2 The same guidance recommends keeping lightweight references and loading details just in time, rather than preloading everything.2 A skill is a tidy version of that idea.
It is hard to maintain
When one document governs your voice, your pricing, your report format and your client rules, changing one thing risks breaking another. Nobody wants to touch it, so it rots.
It does not travel
A mega-prompt lives in whoever's clipboard it is. A skill is a folder: you can version it, hand it to someone else and reuse it in more than one place.
| The mega-prompt | A set of skills |
|---|---|
| One long document, loaded every time | Small folders, loaded when relevant |
| Change one thing, risk breaking five | Edit one skill without touching the rest |
| Lives in a chat or a note | Lives in a folder you can version |
| Hard to hand off | Easy to share with a teammate |
| Bigger as your business grows | Grows by adding shelves, not pages |
What should you turn into a skill?
Start with work that repeats and has a clear right way. A good test is whether you would hand the instructions to a new hire on day one.
- Brand voice and style rules for anything that gets published.
- Report templates: the sections, order and questions your weekly report must answer.
- Review checklists: the claims check before an ad goes live, the proof check before a case study is published.
- Offer facts: what you sell, the price, what is included, what is not, and the exact wording you approve.
- Process steps: how you onboard a client, how you audit a page, how you clean a lead list.
Keep the exploratory thinking in the chat. If you are still writing your processes down, start with turning tribal knowledge into SOPs. Skills are for the parts of your work you want done the same way every time.
How do you build your first skill?
You do not need to be technical to start. Anthropic includes a skill-creator skill that generates the folder structure and formats the SKILL.md file for you.1
- Pick one repeated task you did at least three times this month.
- Write the steps the way you would teach them to a smart new assistant. Include the common mistakes.
- Add examples: one great output and one bad one with a note on why it is bad.
- Attach what it needs: a template, a checklist or a reference file.
- Test it on a real task. Compare to what you would have done yourself.
- Name an owner and a review date so it does not rot.
Write the skill like you are training a person, not like you are programming a machine. The quality of the instructions is the quality of the skill, and the same rules apply as in writing SOPs for agents.
Erik Schluntz of Anthropic gave the same advice for agents generally in a conversation posted on October 17, 2025, one day after Skills launched:
You should still start sort of by from the simplest possible thing and then work up to a more complex solution.
My translation: one skill, one repeated task, tested on real work before you build the second.
VideoAnthropic
Building more effective AI agents
The same conversation covers Skills, sub-agents and tool design from the people building them.
Where are skills available?
According to the announcement:1
- Claude apps: Pro, Max, Team and Enterprise plans. Team and Enterprise admins must enable Skills organization-wide first.
- Claude API: skills can be added to Messages API requests, with versioning through a skills endpoint. On the API they require the Code Execution Tool beta.
- Claude Code: install skills through plugins from a marketplace, or add them to a skills folder in your home directory. The Claude Agent SDK also supports skills.
Anthropic's partner quotes give a flavor of the pitch. Rakuten said of a finance task, "What once took a day, we can now accomplish in an hour."1 That is a vendor-reported anecdote, so take it as the pitch and run your own test.
What is the difference between skills and MCP?
They solve different problems. The Model Context Protocol, introduced by Anthropic in November 2024, is an open way to plug AI tools into the places your data lives.3 That is about access: reaching your files, apps and databases (see Claude Integrations).
Skills are about know-how: how your business does a task once the assistant can reach the material. I would expect many setups to use both. One connects the assistant to your systems, the other teaches it your way of working.
The shift in how to think about it
The old question was: how do I write a better prompt? The new one is: what does my business know how to do, and have I written it down well enough for someone, or something, to do it without me?
That second question is worth answering even if you never touch a skill. Written-down know-how is the real asset. The packaging just makes it usable. Capturing that know-how is the starting point of the systems I build for clients.
Keep building.Ekai
Questions people ask
What are Claude Agent Skills?
Per Anthropic, Skills are folders that include instructions, scripts and resources that Claude can load when needed. They were introduced on October 16, 2025 and work across Claude apps, Claude Code and the API.
How is a skill different from a long prompt?
A long prompt is loaded in full every time. A skill is loaded only for relevant tasks, and Claude pulls in just the minimal information and files it needs. That keeps the working context lean and the instructions easier to maintain.
Who can use Agent Skills?
Anthropic lists availability for Pro, Max, Team and Enterprise plans in the Claude apps, with Team and Enterprise admins enabling them first. Skills can also be added to Messages API requests and installed in Claude Code.
Are Skills safe to install?
Treat them like software. Because skills can include executable code, Anthropic advises choosing them carefully and sticking to trusted sources.
Sources
- 1
- 2
- 3
- 4
