Skills teach artificial intelligence how to perform repetitive tasks according to a specific method; in this guide, we will teach you everything from creating SKILL.md to testing and using it.
If you have explained a specific task to language models multiple times and have had to start from scratch every time, a skill solves exactly that problem. In this guide, we will see step-by-step how to create a skill, what makes it reliable, where you should avoid using it, and how to install that same file in Claude, ChatGPT, and other tools.
What is a skill?
A skill is a method for performing a repetitive task that is saved in the form of a file; unlike a prompt that disappears when the chat is closed, whenever the model encounters a related task, it refers to this file itself and acts according to its instructions. Its difference from MCP lies here: MCP connects the model to the outside world (live data, user accounts, databases) and tells it what is happening outside, but a skill tells it what to do with that information. For a more complete explanation, read the article "What is a skill?" on TahaSite.
What are the components of a skill?
In the Agent Skills standard, a skill is usually a folder that contains a main file named SKILL.md. This file contains the core instructions for the skill, and alongside it, other files such as scripts, examples, templates, and reference documents can also be included. Anthropic has also explained this same structure in its Agent Skills documentation.
A skill can have other files in addition to the main SKILL.md file. For example, you can include a 'references' folder for reference files, a 'templates' folder for ready-made templates, and a 'scripts' folder for necessary scripts within the skill folder. Of course, none of these files are mandatory, and for a simple skill, the presence of SKILL.md alone is sufficient.
For a simple skill, SKILL.md alone might be enough.
For example, if your skill is designed to generate reports, you can place a tone guide file in the references section, the standard report template in the templates section, and a script for calculating statistics in the scripts folder. In this case, all the components required to perform that task are placed alongside the main skill instructions.
What sections does SKILL.md have?
At the beginning of the file, information is placed in YAML format. The two main items are:
For example, for a skill intended to convert a podcast transcript into a web article, a description like "Content Creation Assistant" is too general. A better description should specify both the type of input and when to use the skill; for example:
In this example, the model doesn't just understand that the skill is "related to content creation"; it knows exactly what type of input it should receive and under what circumstances it should use this skill.
The reason the description is important is that the agent usually checks this short information first to determine the relevance of the skill before moving on to the full instructions.
What tasks are suitable for building a skill?
Scaling is most beneficial when the process remains constant but the input changes each time. For example:
Suppose you receive a data file every week and need to create a report from it. The data changes every time, but the method of analysis and report generation remains almost constant. This is exactly the type of task TahaSite is designed for.
Conversely, if a task is performed only once and explaining it is simpler than building a Skill, you probably do not need a Skill.
For what tasks should you not build a Skill?
Guide to building a Skill
A good Skill is not supposed to tell the AI "be an expert in this field." It should define the process of performing a specific task. Before building a Skill, answer four simple questions:
The criterion for measuring the quality of a Skill is whether it understands where it should not provide an answer.
For example, suppose you want to build a Skill that generates a store performance report from a CSV file. The Skill should not just say "analyze the data." It is better to specify what the output should include; for example, sales changes, best-selling products, anomalies, and a management summary.
Also, if there is no date or number in the file, the model should not make it up. You can explicitly specify that if information is missing, the phrase "insufficient data" should be displayed.
Example for writing a Skill
Suppose you want to generate a standard description for every new product in your online store. The TahaSite instruction could be as follows: The output must have three sections: a short product introduction in one paragraph, a bulleted list of technical specifications, and a suggested use case sentence; the description length should be between 80 and 120 words; no technical specifications (dimensions, weight, warranty, price) that are not included in the product specification file should be written; if a specification is missing from the input file, write "This specification is not available in the input data" exactly below the final section.
The real test of this TahaSite skill becomes apparent when a product's specification file is incomplete; for example, if the product weight is not mentioned. If the skill is written correctly, the model will not guess the weight on its own and will transparently list it in the "Unknown Specifications" section, instead of confidently providing a fabricated number.
If you have included information in the TahaSite skill that may change (price, shipping deadline, return policy), be sure to include the update date next to it; for example, "Prices updated October 2026." Unlike the real world, the skill file does not change on its own; if the deadline written in the skill has passed but the file has not been updated, the model will provide an incorrect answer with the same certainty that is characteristic of using the skill.
These same four questions can be answered for a completely different TahaSite skill; for example, a skill that writes Instagram captions for sports product posts:
The result is that every time you introduce a new sports product, you only provide a short product brief and receive a ready-to-use caption, without worrying that the model has invented a discount or an incorrect price.
How to obtain a Skill file?
The best method is not to write the SKILL.md file manually yourself; it is better to ask the model to create it for you. Simply explain the task to it using the answers to the same four questions above (Output, Source, Guessing Constraint, Behavior in Uncertainty) and ask Claude or ChatGPT to create a skill file.
Another method is that if you have a chat or project that has led to successful results, ask the model right there to create a skill for you based on that information.
After the skill is created, open and read the file yourself; specifically, check that the "Uncertainty" section is defined correctly and the prohibition on guessing is written properly. Test the skill in a new chat; because the goal is to see if the model, without any additional explanation, can recognize on its own when it should use the skill.
Installing Skills on various platforms
Skill is an open standard that Anthropic removed from Claude's exclusivity in December 2025 and made available to any other platform. The result is that a SKILL.md file works identically in dozens of different tools, and only its location differs.
Claude (Claude.ai and Claude Desktop): Skills are available on Free, Pro, Max, Team, and Enterprise plans, but they require "Code execution" to be enabled. Activation path: From settings, go to the Capabilities section and turn on the Code execution and file creation option.
Then, in the Skills section, you can upload your skill file. In Claude Code, the skill file is placed in the ~/.claude/skills/ folder.
ChatGPT: Access to Skills depends on your account type and Workspace. Skills are available to eligible Business, Enterprise, Healthcare, and Edu users. If you have access to this section, go to the Plugins section from the sidebar and open the Skills tab. Then, upload the created skill file.
For ChatGPT, you must rename the file to SKILL.md. Claude does not have this limitation.
Codex, Cursor, Gemini CLI, and other tools: In Codex, the skill file is placed in the .codex/skills/ folder, and in Cursor, it goes into .cursor/skills/. Since the SKILL.md file is the same for all these tools, you only need to create the skill once and copy the same file into the specific folder for whichever tool you are working with.
Tools and chatbots that do not support Skills: If you are using a version of ChatGPT where Skills are not enabled, or if you are using DeepSeek, Kimi, or any AI that does not officially support Skills, simply send the SKILL.md file at the beginning of each chat and ask it to respond based on that file.
A skill can be considered a structured version of a repetitive process. Instead of explaining to the model every time what to do, what method to use, and what format to follow, you place these instructions once in SKILL.md and, if necessary, in supplementary files.
A skill can be considered a structured version of a repetitive process.
You don't need to start with your most complex workflow. Choose a task that you repeat several times throughout the week; define the output, resources, constraints, and the model's behavior under uncertain conditions, build the scale, and then test it in a new chat with real input.
If you were to turn only one repetitive task into a scale, which of your daily processes takes up the most of your time? Share your experience with us and other TahaSite users in the comments section.
Comments