If you have ever explained the same task to an AI assistant multiple times, only to start from scratch every time you open a new chat, custom AI skills can help solve that problem.
In this guide, we’ll walk through how to create a skill, what makes it reliable, when you should avoid using one, and how to install the same skill across Claude, ChatGPT, and other compatible AI tools.
What Is an AI Skill?
An AI skill is a reusable set of instructions that defines how to perform a specific task. Instead of repeating the same instructions in every conversation, you save them in a file that an AI assistant can use whenever the task comes up.
Unlike a regular prompt, which is usually tied to a particular conversation, a skill provides a reusable workflow that can be applied across multiple tasks and sessions, provided the platform supports it or the file is supplied manually.
Skills also serve a different purpose from Model Context Protocol (MCP). MCP connects AI assistants to external systems, such as live data sources, databases, and user accounts. A skill, by contrast, explains how the assistant should carry out a task.
In simple terms, MCP helps an AI assistant access information and tools, while a skill provides instructions for using them to complete a particular job.
For a more detailed explanation, see our guide to [What Are AI Skills?].
What Does an AI Skill Contain?
Under the Agent Skills standard, a skill typically consists of a folder containing a main file named SKILL.md. This file holds the core instructions that define the skill. Additional files, such as scripts, examples, templates, and reference documents, can be included when needed.
Anthropic's Agent Skills documentation describes this general structure.
A skill folder might contain the following files:
- SKILL.md: The main instructions that define the skill.
- references/: Reference documents, guidelines, and supporting information.
- templates/: Reusable templates for reports, articles, or other outputs.
- scripts/: Scripts that help automate calculations or other parts of the workflow.
These additional folders are optional. For a simple skill, a single SKILL.md file may be all you need.
For example, if you create a skill for generating business reports, you could include a writing-style guide in the reference folder, a standard report layout in the templates folder, and a script for calculating statistics in the scripts folder.
This approach keeps everything needed for a recurring task organized in one place.
What Should a SKILL.md File Include?
At the beginning of a SKILL.md file, you'll typically find metadata written in YAML format. Two of the most important fields are:
- name: The name of the skill.
- description: A short explanation of what the skill does and, more importantly, when it should be used.
Consider a skill designed to turn podcast transcripts into publishable articles.
A description such as «Content Writing Assistant» is too broad. It doesn't tell the AI exactly what the skill does or when to activate it.
A more useful example would be:
This description tells the AI not only what the skill is about, but also what input it expects and under which circumstances it should be used.
The description matters because an AI agent may initially rely on this short summary to determine whether a skill is relevant before reading its full instructions.
What Tasks Are Best Suited to AI Skills?
Skills are most useful when the process stays relatively consistent while the input changes from one task to another.
Here are some practical examples:
- Turning raw reports into a consistent format.
- Reviewing articles against a predefined checklist.
- Generating financial reports from monthly data files.
- Reviewing code and running a specific set of tests.
- Creating presentations based on company templates.
- Converting meeting notes into decisions and action items.
- Preparing social media content according to brand guidelines.
Imagine that you receive a new data file every week and need to turn it into a report. The data changes each time, but the analysis process and report structure remain largely the same.
That is exactly the kind of recurring workflow a skill is designed to handle.
On the other hand, if you only need to complete a task once and explaining it in a prompt is easier than creating a reusable file, building a skill probably isn't worth the effort.
When Should You Avoid Creating an AI Skill?
Not every task needs a dedicated skill. In some situations, creating one can add unnecessary complexity.
One-time tasks: If you can explain the task in a short prompt and don't expect to repeat it, a regular instruction is usually sufficient.
Information that changes frequently: Avoid hardcoding details such as prices, deadlines, and return policies into a skill. Instead, instruct the AI to check an appropriate, up-to-date source.
Tasks that depend on live information: A skill is not automatically connected to external systems. If a task requires checking the current status of a user account, accessing a live database, or retrieving real-time information, you may need an appropriate integration or tool, such as an MCP connection.
Irreversible actions: Payments, refunds, and data deletion should require appropriate human authorization. A skill can define the workflow, but it should not eliminate safeguards for consequential actions.
How to Create a Reliable AI Skill
A good skill should not simply tell an AI assistant to «act like an expert.» Instead, it should define a clear process for completing a specific task.
Before creating a skill, answer these four questions:
- What exactly should the output look like?
- Which sources should the AI trust when performing the task?
- What information must the AI never guess?
- What should the AI do when information is missing or incomplete?
A reliable skill should define not only how to produce a good answer, but also when not to produce one.
For example, suppose you want to create a skill that generates a store performance report from a CSV file.
Instead of simply instructing the AI to «analyze the data,» specify what the report should include, such as changes in sales, best-selling products, unusual results, and an executive summary.
You should also explain how the AI should handle missing information. If a date or figure is absent from the file, it should not invent one. You could instruct it to display «Insufficient data» whenever the available information doesn't support a conclusion.
These rules make the skill more dependable and reduce the risk of fabricated information.
Example: Writing a Skill for Product Descriptions
Suppose you run an online store and want every new product listing to follow the same format.
Your skill might include these instructions:
- The output must contain three sections: a short product introduction, a bulleted list of technical specifications, and one sentence recommending a suitable use.
- The entire description must contain between 80 and 120 words.
- Do not invent technical details, including dimensions, weight, warranty information, or price.
- Use only the information provided in the product specification file.
- If a required specification is missing, write «This specification is not available in the input data» beneath the final section.
The real test comes when you use the skill with an incomplete product specification file.
For example, if the product's weight is missing, a properly designed skill should not guess the weight. It should clearly identify the missing information rather than confidently supplying an unsupported number.
If your skill contains information that may change, such as prices, shipping deadlines, or return policies, include a last-updated date where appropriate.
For instance, you might write «Pricing last updated: October 2026.»
However, adding a date does not make the information current automatically. You must update the file when the underlying information changes. Otherwise, the AI may continue using outdated instructions or figures.
Another Example: Creating a Social Media Caption Skill
The same four questions can help you design a completely different skill, such as one that writes Instagram captions for sports products.
What should the output look like?
A caption of no more than three lines, written in an energetic, informal tone. It should end with a call to action and include five to seven relevant hashtags related to the sport and brand.
Which sources should the AI use?
Only the product name, intended use, and features provided in the product brief.
What must the AI never guess?
Price, discount percentages, and stock availability. These details can change daily, and inventing them could make the caption inaccurate.
What should happen when information is missing?
If the brief doesn't include a price or discount, the AI should use a predefined sentence instead of inventing a figure.
For example:
«Visit the link in our bio for current pricing and availability.»
Once this skill is ready, you can introduce a new sports product by providing a short product brief. The AI can then generate a caption in the required format without inventing prices or discounts.
How to Have AI Create a Skill for You
You don't necessarily need to write the entire SKILL.md file yourself. In many cases, it's easier to ask an AI assistant to create it.
Start by answering the four questions described earlier: the expected output, trusted sources, restrictions on guessing, and the required behavior when information is missing.
Then ask Claude or ChatGPT to generate a skill file based on your requirements.
For example, you could use this prompt:
Create an AI skill for writing product descriptions from product specification files. The output must contain three sections: 1. A short introduction in one paragraph. 2. A bulleted list of technical specifications. 3. One sentence recommending a suitable use. The complete description must be between 80 and 120 words. Use only information provided in the input file. Do not invent numbers, technical details, or product features. If a specification is missing, write: «This specification is not available in the input data.»
Another approach is to use an existing conversation or project that has already produced consistently good results. Ask the AI to analyze the successful workflow and turn it into a reusable skill.
Once the file has been generated, open it and review the instructions carefully. Pay particular attention to how it handles missing information and whether its restrictions against guessing are clear.
Finally, test the skill in a new conversation. The goal is to verify that the assistant can recognize when the skill is relevant and follow its instructions without requiring you to explain the entire workflow again.
How to Install AI Skills on Different Platforms
Agent Skills is an open standard developed by Anthropic that has been adopted by multiple AI tools. The same general SKILL.md instructions can be reused across compatible platforms, although installation procedures and supported features vary.
Here is how to approach skills on several popular platforms.
Claude (Claude.ai and Claude Desktop)
Claude supports skills on eligible plans and configurations. Availability and setup requirements may vary by product and account.
In supported interfaces, you can typically manage skills through the relevant settings or customization area. Depending on the version, you may need to enable code execution and file creation before using certain skills.
If your interface provides a Skills section, upload your prepared skill file there and follow the platform's instructions.
For Claude Code, skills can be stored in a directory such as:
.claude/skills/
A project-specific skill might be organized like this:
.claude/ └── skills/ └── product-description/ └── SKILL.md
Consult the current Claude documentation for the exact setup supported by your version.
ChatGPT
Skill availability in ChatGPT depends on your account, workspace, and the features enabled for you.
If your workspace provides a Skills feature, open the relevant customization area, locate Skills, and upload the skill file according to the available instructions.
If native skill support is unavailable in your version of ChatGPT, you can still use the instructions manually by attaching or pasting the contents of SKILL.md into a conversation and asking ChatGPT to follow them.
This method is less automatic than a native skill integration, but it can still help you reuse a consistent workflow.
Codex, Cursor, Gemini CLI, and Other Tools
Several AI development tools support reusable skills or similar instruction-based workflows.
For example, Codex and Cursor support skill-oriented directory structures, while other tools may use their own configuration files or conventions.
Depending on the tool, a skill might be stored in a location such as:
.codex/skills/
or:
.cursor/skills/
The precise directory structure and activation behavior depend on the tool and version.
The key advantage of a reusable skill file is that you can maintain one core set of instructions and adapt its installation to each compatible platform rather than rebuilding the workflow from scratch.
AI Tools Without Native Skill Support
What if you use a chatbot that doesn't officially support Agent Skills?
You can often use a simplified manual approach.
Upload SKILL.md at the beginning of a conversation and instruct the AI assistant to follow its instructions for relevant tasks.
For example:
«Read the attached SKILL.md file and follow its instructions whenever I ask you to perform the task it describes. Use only the supplied information, respect all formatting requirements, and do not invent missing details.»
This approach can work with many chatbots, including tools that don't have a dedicated skill installation feature. However, the assistant's ability to follow the file consistently will depend on the platform and its capabilities.
Turn Repetitive AI Tasks Into Reusable Workflows
An AI skill is essentially a structured version of a repeatable process. Instead of explaining what to do, how to do it, and how to format the result every time, you define those instructions once in SKILL.md, adding supporting files when necessary.
You don't need to begin with a complicated workflow. Choose a task you perform several times a week, define its expected output and trusted sources, establish clear limits on what the AI can assume, and specify what should happen when information is missing.
Then create the skill and test it using real input in a fresh conversation.
Over time, a well-designed collection of skills can help you standardize content creation, streamline reporting, improve consistency, and reduce the amount of repetitive prompting required to get useful results.
If you could turn just one recurring task into an AI skill, which workflow would save you the most time? Share your thoughts and experiences in the comments.
