How to Use Claude Skills for Repeatable Workflows

Claude Skills are reusable packages of instructions, scripts, templates, and reference files for a specific kind of work. Claude scans the enabled skills, loads one when the task matches its description, and follows the packaged process.

A skill is useful when you repeatedly explain the same method: how to format a report, review a draft, create a branded presentation, clean a spreadsheet, or turn research notes into a standard brief. It does not automatically remember changing facts or learn a habit in the background; you must update the skill or provide the current data.

What makes a skill different from other Claude features?

• Skill: a task-specific procedure that Claude loads when it is relevant. It can include instructions, examples, resources, and executable code.
• Project: a workspace that keeps related chats and background material together. Project context is useful across the work in that project, not only for one type of task.
• Memory or preferences: user information and preferences available for personalization where that feature is enabled.
• Connector: access to an external service or current data source, subject to the connector's permissions.

These features can work together. A report-writing skill can define the method, a Project can contain the client background, and a connector can retrieve current records. The skill itself should not pretend that a static file contains live data.

Who can use Claude Skills?

Anthropic's current Skills documentation lists support for Free, Pro, Max, Team, and Enterprise users. Skills require Code execution and file creation to be enabled. Organization owners can control availability and provision approved skills for Team or Enterprise workspaces.

Claude also includes maintained skills for tasks such as creating or editing spreadsheets, Word documents, presentations, and PDFs. Feature availability and administrative settings can change, so use the account's current Skills page as the source of truth.

Enable and use a built-in skill

1. For a personal Free, Pro, or Max account, open Settings > Capabilities and enable Code execution and file creation.
2. For a Team or Enterprise account, an owner may need to enable Code execution and Skills in the organization's settings.
3. Open Customize > Skills.
4. Turn on the built-in skill you want to make available.
5. Start a new request that clearly describes the deliverable and supplies the necessary source files.

Claude normally selects a relevant enabled skill automatically. You can be explicit while testing: “Use the presentation skill to create a six-slide project update from the attached status report.”

After Claude creates a file, inspect its content and open it in the destination application. A successful download does not prove that formulas, page breaks, citations, charts, or accessibility are correct.

Choose a good task for a custom skill

The best first skill has a stable input, repeatable steps, and a clear output. Good candidates include:

• turning meeting notes into an agreed decision-and-action format;
• reviewing an article against a publication checklist;
• converting a monthly CSV export into the same cleaned workbook;
• creating client slides with an approved template and brand rules;
• extracting claims that require source verification; or
• producing a standard research brief from a defined set of documents.

Avoid a skill such as “do all my marketing.” It has no precise trigger, input contract, or definition of done. Split it into smaller skills such as campaign-brief review, social-post formatting, and launch-report generation.

What a custom skill should contain

A custom skill is packaged as a folder. Its central file is SKILL.md, which describes the skill and tells Claude what to do. The folder can also contain templates, examples, scripts, and reference material.

A minimal structure might look like:

weekly-status/
  SKILL.md
  references/
    status-template.md
    approved-terms.md

The instructions should define:

• when the skill should and should not be used;
• required inputs and what to do if one is missing;
• the steps in the correct order;
• the required output structure and file type;
• validation checks before completion;
• facts that must be cited or left unresolved; and
• actions that require human approval.

Example instruction

# Weekly project status

Use this skill when the user asks to turn project notes into a weekly status update.

Required inputs:
- reporting period
- project name
- source notes or tracker export

Process:
1. Use only the supplied sources.
2. Separate completed work, current work, risks, decisions, and next actions.
3. Give every action an owner and due date when the source provides them.
4. Mark absent owners or dates as "Not provided"; do not infer them.
5. Flag numerical claims that do not have a traceable source.
6. Follow references/status-template.md.

Before delivery:
- confirm every named person and date appears in a source;
- remove duplicate actions;
- preserve unresolved disagreements;
- do not send or publish the report.

This is more reliable than vague language such as “make the status professional.” It specifies both the workflow and the boundaries.

Create and upload a custom skill

1. Create the skill folder and required SKILL.md file. Add only the supporting resources the workflow needs.
2. Compress the folder as a ZIP file.
3. Open Customize > Skills in Claude.
4. Select the add button, choose Create skill, and then Upload a skill.
5. Upload the ZIP and enable the new skill.
6. Test it with a normal request, a request that should not trigger it, and a request with a required input missing.

Anthropic also provides tools for creating skills interactively, and some plans and surfaces can record a workflow for Claude to turn into a proposed skill. Recorded workflows can capture everything displayed or spoken during the session, so close private material and never enter passwords or secrets while recording.

How to test a skill

Do not test only the happy path. Use a small test set:

1. Direct match: a typical request that should activate the skill.
2. Paraphrased match: the same intent in different wording.
3. Near miss: a related task that should not use the skill.
4. Missing input: confirm Claude asks for or clearly marks required information instead of inventing it.
5. Conflicting sources: verify that the output preserves the conflict.
6. Badly formatted input: test a realistic messy document or spreadsheet.
7. Output validation: open the resulting file and check every required field, formula, citation, and format rule.

If Claude does not use the skill, confirm that it is enabled and improve the skill description so the trigger is specific. During testing, explicitly name the skill. If it triggers too often, narrow the description and add a “do not use for” section.

Practical skill ideas

Research synthesis

Define an evidence order, required citation format, treatment of contradictory sources, and a rule that unsupported claims stay unresolved. A skill can organize research; it cannot make a weak source true.

Factual-review checklist

Ask the skill to extract dates, figures, names, product capabilities, and causal claims into a verification list. Require a source beside each claim and prohibit changing “unverified” to “verified” merely because the wording sounds plausible.

Meal planning

Store stable household preferences and a reusable output format, but provide current pantry items, schedule, budget, and dietary changes each time. Keep allergy and medical nutrition decisions under qualified human oversight.

Document or slide production

Bundle an approved template, typography rules, color values, image-placement guidance, and a preflight checklist. Test the final file visually; a template does not prevent overflow, clipped charts, or inaccessible contrast.

Spreadsheet cleanup

Specify source columns, data types, duplicate rules, calculations, error handling, and a reconciliation total. Require formulas to be preserved and the original data to remain on a separate sheet.

Security risks when installing a skill

Skills can include scripts, dependencies, and instructions that use network access. Anthropic identifies prompt injection and data exfiltration as important risks. A skill shared by a colleague is not automatically safe.

• Install skills only from sources you trust.
• Read every instruction, script, dependency, and bundled file before enabling it.
• Look for commands that upload data, access secrets, install packages, or contact unfamiliar domains.
• Give the code-execution environment only the network access it needs.
• Test with non-sensitive sample data first.
• Keep sending, publishing, paying, deleting, and account changes behind human approval.
• Version and review organization skills like other operational code.

The official usage guide also explains organization provisioning, sharing controls, and troubleshooting. On Team and Enterprise plans, skill sharing can be disabled by default or limited by administrators.

Maintain the skill

Assign an owner, version, and review date. Update the skill when the business process, template, policy, software interface, or source-of-truth location changes. Retest after each material change with the same cases used for the first version.

Track whether the skill actually reduces corrections and time. If users routinely override a step, the documented workflow may be wrong. If the skill keeps growing, split it into focused components that Claude can combine when needed.

Claude Skills work best as executable standard operating procedures: narrow enough to trigger correctly, explicit enough to test, and bounded enough to review. They improve consistency, but the user remains responsible for current data, factual validation, permissions, and approval of consequential actions.