A SKILL.md file is an entry point for putting a workflow into a compatible AI application. It describes when the skill applies and how to do the work; its folder can hold supporting references, templates, examples or scripts. The workflow might prepare a project update, review an intake or draft a customer handoff. (Agent Skills specification)

Before writing one, define the task, its inputs and the standard a reviewer expects. Those choices give the instructions a practical purpose.

A skill carries the method from one task to the next.

A prompt is a request. A programmatic AI workflow is a procedure built for repeatable use: a trigger, inputs, steps, templates, checks, exception rules and an output. A person can start it; a scheduled run needs an application configured to start it.

Shared instructions and review criteria guide the work. People set the standard, review outputs and decide what happens next. The application needs suitable tools, permissions and testing. AI prepares; people decide. Illustration: one capability may use several workflows: a closed-deal handoff can prepare an onboarding checklist, billing setup draft and missing-information list.

Illustration: a weekly project brief.

Consider an invented professional-services team preparing a weekly project brief. A request to summarize the week's notes leaves several choices open: which dates count, how to flag a blocked decision, and whether a suggested change is already approved. A skill could specify the sources to read, separate confirmed commitments from proposals, and produce a draft with unresolved questions for the project lead.

The aim is a shared method colleagues can inspect and improve, without reconstructing the instructions from an old conversation. The person requesting the work still reviews the draft and resolves its open questions.

The file is the entry point to a small package.

Under the open Agent Skills specification, the file begins with a metadata header in YAML, a text format for named fields. It must include a name and description. The instructions follow in Markdown: plain text with notation for headings, lists and links. Optional supporting files can hold longer references, reusable templates or code for particular operations. (Format specification)

The handover package gives the team the material to run, check and maintain the work: instructions, templates, references, exception rules, test inputs and expected results, a record of performance on tested cases, and handover notes naming an owner. Agree where the maintained package lives and who can update it.

For a skill-based implementation, the skill folder is the part installed in a compatible application. Evaluation assets can be maintained alongside it. Installation and configuration vary by platform; the skill format does not require a particular evaluation folder or automatically run its tests. Keep reference answers outside the tested runtime's access.

Illustration: a skill folder.

The following folder layout is an illustration, not a tested download. It keeps the installed skill and its evaluation material separate within one maintained release.

workflow-release/
  weekly-project-brief/
    SKILL.md
    references/
      project-status-rules.md
    assets/
      briefing-template.md
  evaluation/
    cases/
    reference-results/
    review-criteria.md
    run-log.md

Keep required supporting files with the installed skill. Maintain evaluation cases and reference outcomes, sometimes called golden records, separately. Reserve fresh cases from development; during testing, supply their inputs while keeping expected answers unavailable to the tested runtime, including its file and tool access. A sibling folder alone does not enforce that separation. Use invented or permitted material. A plain-language workflow may need no scripts. The testing guide explains how to judge the result.

Skills can support work people start and review.

Anthropic documents skills for creating documents and spreadsheets, applying brand guidance and following company-specific procedures. A person can request the work, review the output and decide what happens next. (Claude skills)

The company decides who can use the material, which sources are authoritative and who approves the work. Write those responsibilities into the workflow alongside the preparation steps.

Start with existing know-how or a capability you need.

For an established job, begin with the person who knows it. Their working examples, exceptions and explanations may be more useful than an old SOP alone. Capture what they actually check and the questions they ask when information is incomplete.

For a new function, start with a selected method and a responsible owner. Agree what a useful output should contain, adapt the method to the business and test the result with the people who will use it.

Both paths need business context and a way to judge the work. A file downloaded from a library supplies a starting point; the team needs to examine its instructions, dependencies and suitability for the job.

Check the workspace as well as the instructions.

Check required tools, file types, connections and permissions in the application where the team will use the skill. The file supplies instructions; the application controls access and execution. Try representative tasks in that workspace before sharing it more widely. (Compatibility in the specification, application integration guide)

An AI application can draft a package, but the team still needs to review its contents and test what it can do. Instructions express the intended behavior; permissions and approval controls must be enforced by the surrounding software. Read the security guide before installing unfamiliar packages or connecting business systems.

At Fully Powered AI, we capture existing know-how through Fully Powered Studio and help teams establish missing capabilities from reusable workflow foundations. We adapt the instructions and context, test the outputs and help the team put them into use.

Choose one piece of work your team wants to do well more consistently. The SOP-to-workflow walkthrough shows what to specify before building the skill.