What an AI skill actually adds to your assistant

Imagine helping a community garden write its weekly newsletter. Every Friday, someone supplies a handful of notes. Every Friday, the instructions are much the same: put the next event first, keep the tone welcoming, include only confirmed details, and flag missing information before anyone sends it.

You could paste those rules into each conversation with an AI assistant. An agent skill gives you a way to package that routine, its template and any supporting material so a compatible assistant can find and reuse it.

This week, Anthropic announced on October 22 that it was donating the Agent Skills standard to the Agentic AI Foundation. The foundation’s announcement confirms the addition. The format already existed; the news concerns its home and governance. For ordinary users, it is a useful moment to understand what is actually being shared.

A recipe the assistant can read

Under the Agent Skills specification, a skill is a folder with a file named SKILL.md. The file contains a name, a description of when the skill is useful, and instructions. The folder can also hold reference documents, templates or scripts.

Think of a recipe folder: a label helps you find the dish, the recipe explains the steps, and an extra page might explain a technique. For the garden newsletter, the recipe could specify how to turn rough notes into a draft. A template could show the sections. An example could show the desired tone.

This mechanism supplies material for the assistant to read and use. Loading it is not the same operation as retraining the model. The standard’s overview describes a staged process: first discover the skill’s name and description, then load its instructions when relevant, then consult extra resources as needed.

That arrangement explains why the description matters. “Helps with writing” gives little guidance. “Draft the garden’s weekly newsletter from supplied event notes” identifies a particular job. The instructions can stay detailed without every unrelated conversation needing the whole folder.

What would our newsletter routine contain?

Here is an invented brief for the garden. It is an example of the decisions a skill could capture, rather than a report of a working installation:

• Input: use the current week’s supplied notes and the newsletter template.
• Order: lead with the next confirmed event, then list other updates.
• Missing details: put unanswered questions below the draft; do not fill gaps with plausible guesses.
• Output: prepare text for the organizer to review. Sending is a separate step.

Suppose the notes say that Saturday’s seed swap starts at 10 a.m., while a compost workshop is proposed but has no date. A suitable draft would lead with the seed swap. It would leave the workshop out of confirmed events and ask the organizer for its status. The newsletter can still move forward without pretending the missing date is known.

Now add a style example: short paragraphs, descriptive headings and no invented quotes from volunteers. These are choices someone can inspect and edit in the routine. If the newsletter changes next month, the instructions can change with it.

Reusable does not mean identical everywhere

A shared format helps different products recognize the same kind of package. It does not make their environments identical. Microsoft’s VS Code documentation explicitly notes that skills are portable across compatible agents while supported locations and optional features vary.

Our newsletter routine might work with notes supplied directly in a conversation. A version that fetches the garden’s calendar needs a way to access that calendar. A version that creates a designed document needs the appropriate document tools. A sentence in the instructions cannot supply a missing connection or executable.

The specification allows authors to state compatibility requirements. Read them before assuming a downloaded folder will work in your assistant. Also check how that product discovers or invokes skills. Having a skill available is not proof that it was used for a particular answer.

Inspect the recipe, then try awkward ingredients

Skills can include executable scripts, so their contents deserve attention beyond the title. Microsoft’s guidance recommends reviewing shared skills and describes separate controls over script execution. For a newsletter routine, inspect what files it reads, whether it calls an outside service, and whether it tries to send anything. Keep the implementation proportionate to the job.

Then try a few small cases before relying on it. For our invented garden, a useful trial set would include:

1. Complete notes: does the draft preserve the supplied date, time and location?
2. A missing time: does it ask for the detail instead of inventing one?
3. A cancellation: does it avoid advertising the cancelled event as upcoming?

Keep the inputs and outputs so you can compare what happens after an edit. If the assistant ignores a rule, first check whether it loaded the intended skill and had the required information. If it did, the procedure or the model’s execution may still need work. A saved routine gives you something concrete to improve; it does not certify the resulting text.

The most useful first skill is often a task whose rules you can already explain. For the garden, success would mean spending less effort repeating the same instructions while keeping the organizer’s judgment visible: what counts as confirmed, what belongs in the draft, and what still needs an answer.