How to create a Claude Skill for repeatable work

Start with a process worth repeating

A Claude Skill is a reusable set of instructions and supporting material that Claude can use when a task matches it. Anthropic describes custom skills for specialized knowledge and workflows. They can be simple Markdown instructions; scripts are optional.

Choose a small process you can judge: turning approved meeting notes into an internal action brief, for example. The skill can define headings and checks, but it cannot know whether the notes are true. Keep the source facts in the individual request. Do not package customer records, live passwords or a confidential document as a general example just because they are convenient.

Define the trigger, inputs and finish line

Before creating a skill, write down when it should be used and when it should not. For the action brief, the trigger is a request to turn meeting notes into a short internal follow-up. The inputs are the approved notes and the intended reader. The finish line is a list of decisions, actions with recorded owners and dates, and open questions. If the notes omit an owner, the skill must leave it unassigned.

Anthropic recommends a specific description, focused instructions and examples. A skill named General Office Helper is too broad: Claude cannot tell which jobs belong to it. A name such as Meeting Follow-up Brief is easier to select and test. Keep a human owner for reviewing facts before the output is circulated.

Check access before building

Anthropic currently lists Skills on Free, Pro, Max, Team and Enterprise plans, with code execution and file creation enabled. Individual users find skills under Customize > Skills; personal accounts may first need the capability enabled in Settings. Team and Enterprise owners can control availability in organisation settings. If you cannot see the option, check your current help page and ask your workspace owner rather than assuming the skill was uploaded incorrectly.

The same help centre also describes built-in skills for documents and spreadsheets. Try an existing capability before creating another process to maintain.

Write a first version without code

Anthropic's creation guide describes a skill folder with a Markdown instruction file containing a name and description. The body can explain the workflow and show examples. Package the folder as a ZIP for upload through Customize > Skills, using the current official instructions for file naming and structure. Read the final package before upload; a skill can later contain executable files, so treat downloaded skills like software from an outside source.

Keep version one text-only. Include what input is required, the output headings, how to mark missing information, and a short review checklist. Do not tell Claude to send the brief, update a CRM or contact a client. Those are separate actions needing their own authority and checks.

Help me draft a text-only Claude Skill called Meeting Follow-up Brief. It should apply only when I ask to turn approved meeting notes into an internal action brief. Required input: notes and intended reader. Output headings: Decisions made, Actions, Open questions. For each action, copy only an owner and due date that appear in the notes; otherwise write not recorded. Do not invent agreement, send messages, update systems or include customer data in examples. Include two fictional examples and a checklist for a human to compare every decision, owner and date with the notes. Show me the proposed files for review before I upload anything.

Upload and test it with three different notes

Anthropic's current guide says to create or upload a skill from Customize > Skills and toggle it on. After upload, ask Claude naturally for the task the skill covers. Do not assume the skill was used just because the answer looks tidy; inspect the available skill indicators or ask how it handled the instructions, then check the result against your source notes.

Run three fictional tests. First, give complete notes with one decision and two actions; check that all details survive. Second, omit an owner and a date; check that both stay visibly missing. Third, give an unrelated request, such as a holiday itinerary; the meeting skill should not be applied. Fix the description or rules if it triggers too broadly, misses the task, or fills gaps with invented facts. Anthropic recommends testing after each meaningful change.

Share a reviewed skill, not an unfinished experiment

On Team and Enterprise plans, Anthropic documents ways to share a skill with colleagues or submit it to an organisation library, subject to workspace policy and review. A personal upload is not automatically visible to everyone. Ask who owns the process, who can edit the skill, and where the approved version lives. If a team template changes, revise and test the skill before publishing the new version.

Sharing instructions can spread mistakes as efficiently as good practice. Have the process owner check its examples for private information and its rules against the current policy. Provide a short change note so colleagues know what to retest. Recipients should still compare outputs with the underlying notes.

Keep the skill inside its limits

A skill is an instruction package, not an approval, access grant or accuracy guarantee. If it calls tools or contains code, review those components and the permissions they need before enabling it. Anthropic specifically advises caution with scripts, third-party skills and secrets. Start with a text-only process when that is enough.

When the workflow changes, update the skill's source and repeat the three tests. If the output will be sent outside the team or used for an important decision, assign a person to check the final facts and take the action. For help designing a repeatable workflow around your team's real work, use the existing Clearer With AI workshop enquiry.

Common questions

Do I need to write code to make a Claude Skill?

No. Anthropic says a simple skill can use Markdown instructions. Executable scripts are optional and need more careful review.

Will Claude use my Skill in every chat?

No. Anthropic describes selecting a skill when the request matches its description and the skill is enabled. Test both matching and unrelated requests rather than assuming perfect selection.

Can my colleagues see a skill I upload?

A personal upload is not automatically shared. Team and Enterprise sharing or organisation publishing depends on workspace settings and, in some cases, owner review.