The owner's skill library, part 1: turn repeat work into AI skills
What an AI skill is, how to pick your first three, a one-page template you can copy, and who owns, reviews and limits each one.
If you use an AI assistant at work, you may have typed the same explanation more than once: how your proposals are laid out, the tone your customer emails take, where the action items go in meeting notes. Then, in the next chat, you explain it again.
A skill lets you write that explanation down once. Anthropic's Claude Docs put the difference plainly: unlike prompts, skills "load on demand, so you don't have to repeat the same guidance across conversations." Anthropic's engineering team compares building one to "putting together an onboarding guide for a new hire." This series builds a small library of them for your business.
What a skill is
In Anthropic's Claude, a skill is a folder. Anthropic's engineering blog says that "at its simplest, a skill is a directory that contains a SKILL.md file," and that the file starts with two required fields: a name and a description. The instructions go below those fields, written in Markdown. Anthropic's announcement says a skill folder can include "instructions, scripts, and resources that Claude can load when needed."
The name and description matter more than they seem to. Anthropic's engineering team says Claude uses them "when deciding whether to trigger the skill," and Anthropic's announcement adds that "Claude will only access a skill when it's relevant to the task at hand." Claude reads the full instructions only when the task calls for them.
The Claude Help Center separates skills from two related features. Custom instructions "apply broadly to all your conversations," while skills "are task-specific and only load when relevant." MCP (Model Context Protocol) connects Claude to external services and data sources, while skills "provide procedural knowledge": the how-to for a particular task.
You do not need to be a programmer: the Claude Help Center says anyone can create skills by writing instructions in Markdown, with no coding for simple ones. The format is also published as an open standard at agentskills.io, so, in the Help Center's words, "skills you create aren't locked to Claude."
If you would rather not edit files at all, Anthropic describes a "skill-creator" skill that gives interactive guidance: "Claude asks about your workflow, generates the folder structure, formats the SKILL.md file, and bundles the resources you need."
For the case for starting with one agent and a few skills, see our earlier guide, One AI agent with skills often beats a team of bots. Start there..
Pick your first three skills
We suggest starting small. Look for a task that passes four tests:
- It comes up every week.
- It follows the same steps each time.
- You already know what a good result looks like, and you can show an example.
- A person would catch a mistake before it reaches a customer.
Three of the custom skill examples the Claude Help Center lists fit everyday small-business work: applying brand style guidelines to documents and presentations, generating communications that follow company email templates, and structuring meeting notes in a company-specific format. Any of them can be a first skill.
Keep each skill to one job. The Claude Help Center advises creating "separate skills for different purposes rather than a single skill that's meant to do everything," and says the best skills are "focused on one workflow." Anthropic's engineering team suggests a similar order of work: watch where the assistant struggles on real tasks, "then build skills incrementally to address these shortcomings."
Start with three, not thirty. With three, you can test each skill properly and notice when one needs fixing.
The one-page skill template
Copy the block below. The top part is the SKILL.md file itself. The Claude Help Center caps the name at 64 characters and the description at 200, and says the description should cover "what the skill does and when to use it." It also suggests example inputs and outputs "to help Claude understand what success looks like," which go under "What good looks like."
The bottom part is the library card. It stays in your own register, not in the skill.
--- name: ____ description: ____ --- # ____ ## When to use this skill ____ ## When not to use it ____ ## Inputs it needs - ____ - ____ ## Steps 1. ____ 2. ____ 3. ____ ## What good looks like Example input: ____ Example output: ____ ## Rules Always: ____ Never: ____ ## Data it may use ____ ## Data it must never use ____ ## Hand back to a person when ____ ================================================== LIBRARY CARD Keep this in the owner's register, not in the skill. ================================================== Skill owner (one named person): ____ Reviewer: ____ Version: ____ Last reviewed: ____ Next review: ____ Test prompts: 1. ____ 2. ____ 3. ____ Approved data sources: ____ Changes log: ____
Who owns and reviews each skill
Every skill needs one named owner: a person, not a department, who keeps it accurate when the way you work changes. A second person reviews any change before it goes live. Both names go on the library card.
Test after every change. The Claude Help Center recommends testing a skill "with a few different prompts to ensure it works as expected," and testing "after each significant change rather than building a complex skill all at once." Keep three test prompts on the library card so you run the same checks each time.
Ownership matters more once a skill is shared. According to the Claude Help Center, on Team and Enterprise plans shared skills are view-only and recipients automatically get the updated version at next use, and a published skill is managed by your organization, so if your organization requires review, each new version goes through the same review. One edit to a shared skill reaches everyone it is shared with, so one person should be accountable for it.
If a skill no longer matches how you work, switch it off. The Claude Help Center notes that any skill can be toggled off and that "disabled skills won't be available to Claude." Keep it off until the owner updates it and the reviewer signs off.
Data limits
A skill is a set of instructions, and instructions can point at data. Set the limits before you write the steps.
- 01
Keep secrets out.
The Claude Help Center says: "Don't hardcode sensitive information (API keys, passwords)."
- 02
Name the data.
Fill in "Data it may use" and "Data it must never use" in every skill, and list approved sources on the library card.
- 03
Install only from trusted sources.
Anthropic's Claude Docs say to use skills "only from trusted sources: those you created yourself or obtained from Anthropic." The Claude Help Center says to review a skill from a less-trusted source before you enable it, even one shared by a colleague.
- 04
Be careful with skills that reach the web.
Anthropic's Claude Docs warn that skills "that fetch data from external URLs pose particular risk, as fetched content might contain malicious instructions."
- 05
Know the main risks.
The Claude Help Center names prompt injection and data exfiltration as the most significant risks.
- 06
Treat it like software.
Anthropic notes that skills give Claude access to execute code, and its Claude Docs advise extra care when a skill touches systems with sensitive data or critical operations.
Part 2 of this series follows.
Sources
- Anthropic, Introducing Agent Skills, October 16, 2025, updated December 18, 2025
- Anthropic Engineering, Equipping agents for the real world with Agent Skills, Barry Zhang, Keith Lazuka and Mahesh Murag, October 16, 2025
- Claude Help Center, What are skills?
- Claude Help Center, Use skills in Claude
- Claude Help Center, How to create custom skills
- Anthropic, Claude Docs, Agent Skills
- Consultiply, One AI agent with skills often beats a team of bots. Start there., October 26, 2026
Reach us at transform@consultiply.ai
Want this working in your business?
Tell us what you are trying to fix and we will point you at the smallest useful next step.
Talk to us