Claude Skills: enable a reusable workflow and check that it applies

A reusable procedure needs to be available as a native skill before Claude can select it for a task. Open Customize > Skills, check that the intended skill is enabled, and make a small request that clearly matches its purpose. If it is your own skill, inspect its package and description before diagnosing the output.

For example, a note-review procedure can require every proposed action to retain its owner and leave missing dates unresolved. The test is whether the procedure is applied and those facts survive, not whether the answer sounds polished.

Check the capability and the skill toggle

Claude Skills require code execution. For individual accounts, check Settings > Capabilities; managed accounts may need an owner to enable the capability and Skills in organization policy. Then open Customize > Skills and enable the relevant entry. Account and organization controls determine which options you see. Anthropic's Skills instructions

Start with a skill whose purpose you understand. For a downloaded package, read its instructions, scripts, dependencies, and network requests before enabling it. A familiar filename does not tell you what bundled code will do.

Give a custom skill a precise trigger

Anthropic's custom-skill documentation requires a skill directory with a skill.md file beginning with YAML name and description fields. Claude uses the description to decide when the skill applies. Custom-skill structure

Here is an original, instruction-only starting example. It is not an installed or executed skill:

---
name: handover-note-review
description: Review internal handover notes for explicit actions, named owners, and unresolved dates before drafting a short handover.
---

# Handover note review

Use the supplied notes only. Do not invent people, dates, approvals,
or completed work.

First list each proposed action with its owner, deadline, and source
wording. Write Not supplied for a missing owner or deadline.
Keep planned actions separate from completed work.

Then draft a short handover using those supported facts.
End with the specific missing information the author must confirm.
Do not send the handover or contact anyone.

The description names both the material and the moment of use. “Help with work” would leave it competing with many unrelated tasks.

Keep this first version small enough to inspect. You can see the intended evidence boundary and output sequence without reading a collection of scripts. Add supporting resources only when the workflow needs them.

Package and upload your own skill

Place the file in a folder named for the skill, then ZIP that folder. The archive should contain the skill folder with skill.md inside it, rather than loose files at the ZIP root. Anthropic's package structure

handover-note-review.zip
  handover-note-review/
    skill.md

In Customize > Skills, use the add control, choose Create skill, then Upload a skill, and select the ZIP. Enable the resulting entry. If creation or upload is unavailable in a managed account, check the organization's creation policy or ask its owner. Upload and access controls

After upload, compare the entry's name and description with the file you intended to add. Accidentally uploading an older package can look like an instruction-following problem later.

Test with a gap the workflow should expose

Give Claude this fictional request:

Use my handover-note-review skill for these fictional notes:
Mara will check the equipment list. No deadline has been supplied.
The access review is planned; it has not been completed.
The shift lead must confirm who owns the access review.

Prepare the review and a short handover. Do not invent missing values.

Look for three concrete results: Mara remains the equipment-list owner, her deadline stays missing, and the planned access review does not become completed work. The owner of that review should remain unresolved.

Inspect any visible skill-use activity or explanation in the interface as well as the answer. An output resembling the requested format alone does not prove which skill was loaded. Anthropic recommends testing triggering prompts and revising the description when invocation is unreliable. Test a custom skill

Fix availability before rewriting the procedure

If the skill does not apply, check that it is toggled on, its description matches the task, and code execution is enabled. Try naming the skill explicitly in the request. If upload fails, inspect the required file, folder naming, and archive structure before changing the review rules. Skills troubleshooting

For material that belongs together across conversations, Claude Projects address a separate need: reusable context. Keep the skill focused on how to perform the review, and supply the current notes with the actual task.

After a change, repeat the small test and read the resulting handover. Keep the version whose behavior you checked; leave an unanswered owner or date visible until someone supplies it.

Keep reading