Anthropic has unveiled a new method for teaching practical skills to its artificial intelligence models: skills.
Imagine ChatGPT GPTs or Gemini Gems capable of running scripts and mobilizing autonomously to better respond to a user request. Well that’s exactly what Anthropic has been offering since October 16 with the skills.
A concrete example to better visualize the thing? When a user requests the creation of a multilingual website, Claude automatically combines several specialized skills: a first skill generates the code respecting precise guidelines, a second analyzes it and corrects it using scripts, a third writes the content respecting a charter, a fourth translates it with a tone adapted to each country, all throughout the conversation and in total autonomy. The goal is to have a range of skills dedicated to specific tasks that Claude can activate on the fly. If all this makes you think of an AI agent, don’t be surprised, that’s normal: that’s what skills are, a hybrid between the bot and the agent, the difference being played out in the orchestration, here a simple conversation.
In form, each skill consists of a folder and several sub-files (detailed below), directly integrated into Claude’s interface. Once activated, the skill can be invoked by the model when it considers it relevant to respond to the user’s request. Skills can be used in Claude’s desktop and mobile applications as well as via the Anthropic API.
Very concretely, the skills are presented in the form of an archive (.skill extension). The latter contains a markdown file named SKILL.md. The file contains information about the skill name, a description, and the detailed natural language instructions needed to guide the AI. Optionally, the skill can contain additional resources, such as a knowledge base, which must be in markdown format, preferably (.md). Finally, the Skill can also contain scripts in Python or JavaScript (or Node.js). Everything must be coherent and adapted to the task you wish to teach Claude.
How to create your skill?
It is possible to create a skill by hand, by editing the configuration files yourself, or to have it generated by Claude using… a skill creator skill in Claude.
Possibility 1: create a skill manually
The principle of Skill is based around SKILL.md, the main file. The file must start with your skill metadata in YAML format: name for the name, description for the description in natural language of the task that your skill must solve, version for the version number of your SKILL.md, dependencies for the dependencies of your scripts.
The file must then contain, at a minimum, in natural language, the context and objectives of the Skill, the instructions to follow (to guide the AI), the specific use case(s) where Claude must activate it.
Example of a minimal SKILL.MD file:
--- name: Résumeur de texte description: Résume des documents longs en un texte clair et concis version: 1.0.0 dependencies: --- ## Objectif Ce Skill permet à Claude de générer des résumés fidèles et synthétiques à partir de documents longs (articles, rapports, essais, etc.). ## Instructions - Lire l'intégralité du texte avant de commencer le résumé. - Préserver les informations principales, les faits et les conclusions. - Éviter les détails secondaires, les exemples redondants et les citations longues. - Utiliser un style neutre, fluide et compréhensible pour un lecteur non spécialiste. ## Exemple Les exemples sont situés dans : resources/ ## Quand l'utiliser - Lorsqu'un texte dépasse une page ou contient plusieurs sections. - Quand l'utilisateur demande un résumé ou une synthèse.
You can then add examples in a subfolder resources and any scripts in a folder scripts. The tree structure must be clear and the instructions simple to understand. A skill generally allows you to perform a few actions. For more complex schemes, Anthropic recommends creating several skills.
Example of a correct tree structure:
resumeur-de-texte.zip
└── resumeur-de-texte/
├── SKILL.md
├── resources/
│ ├── exemple_texte_resume.md
└── scripts/
Once your skill is complete, rename the file by replacing its extension with .skill. You will then have to import it into Claude to use it (read below).
Possibility 2: create a skill with Claude
To simplify the creation of skills, Anthropic has developed a skill creator skill. Just give them as much information as possible about the purpose of your skill and how to proceed. To proceed, go to the Claude application (Desktop or web). Then click on your profile, on “Settings” then “Capabilities”. At the bottom of the Capabilities page, enable “skill-creator”. Then return to Claude’s main interface and ask the AI to create your skill.
For the example, we are creating a skill at the start of the week, which will: consult our Google Meet calendar to identify appointments for the next 5 days; create a brief reminder of the main meetings; Generate a draft appointment reminder for our various contacts.
Create a skill that allows Claude to check all events in Google Calendar for the next 5 days from today’s date, then generate a structured report chronologically listing each meeting with its date, time, title, participants and videoconference link if available. Then, for each identified meeting, automatically create a draft reminder email in Gmail intended for all participants (except myself) with a clear subject and a courteous message including the exact date and time, the subject of the meeting, the videoconferencing link (Google Meet, Zoom, etc.) and a polite phrase. The skill must present a final summary indicating the number of drafts created and their status, without sending emails automatically, excluding “all day” or events without participants, and correctly managing time zones.
Claude will then call “skill-creator” which will generate each file contained in the final archive one by one.
The “.skill” archive is a compressed container, similar to a ZIP file, bringing together all the components necessary to define a skill for Claude. To use it, simply rename the extension from “.skill” to “.zip”, thus allowing its contents to be decompressed, like a classic folder.
The files generated in our “reminders-meetings” skill by Claude are available for consultation here.
How to use and import a skill?
Before using a skill, you will need to import it into Claude. To do this, once in the Claude application, go to Settings -> Capabilities -> skills. Then click on “Upload skill.” Once the skill has been uploaded, it is directly active. Claude can then detect and spontaneously invoke the most relevant skill based on the implicit needs of the conversation. The model dynamically evaluates the user’s request and automatically selects the appropriate skill, without explicit intervention.
In our case, the skill runs perfectly. A brief reminder of the main meetings is generated and email drafts for our correspondents are written.
Even more interesting for developers: it is now possible to use a Skill directly from the Anthropic API. To do this, simply send the Skill first via a POST request to https://api.anthropic.com/v1/skills, including the Authorization: Bearer YOUR_API_KEY and Content-Type: multipart/form-data headers. Once the Skill has been uploaded, it can be invoked in a request to the model by simply entering its identifier (skill_id) in the request parameters.
When to use a skill?
Anthropic recommends creating a skill when you have already identified a way of working that you want to automate and reuse without re-explaining everything each time in a prompt. In other words, when you notice that you often repeat the same prompts, structures or quality criteria in your exchanges with Claude, it becomes relevant to make it a skill. This saves you from having to re-explain everything with each new request: Claude automatically recognizes the context and applies your approach, your standards and your reference models.
Skills are particularly useful for defining the guidelines of a group: branding, charter, tacit rules, etc. The more you enrich your skills library, the more Claude becomes able to anticipate your needs and act independently and according to your criteria. AI no longer just executes your requests, but truly learns from the way you work.
