Running Claude Skills Without Desktop

Run Claude Skills in Claude on the web for browser-based staff, Claude Code for repository work, or the Messages API for application and CI requests. For a customer-facing link, put the chosen runtime behind your own web app and test the customer journey before sharing it.

Test access from the browser account that needs it

For staff who already work in Claude on the web, start with a browser trial. For an Enterprise pilot, have an owner check Code execution and file creation and Skills in the Policy tab under organization settings. Once the organization enables skills, confirm the intended user can find the provisioned skill or upload their custom skill under Customize > Skills.

Run a representative prompt and check the artifact or response against one rule specific to the skill. Record visibility and output as separate checks. If the owner can see the skill and a teammate cannot, inspect the teammate’s access first. If both see it and produce the same incorrect result, revise the skill instructions and rerun the fixture.

A green process exit can accompany a missing export. Record a small expected output manifest before the trial and compare it with the actual paths before accepting the run.

Use skills in Claude | Claude Help Center

Use an explicit invocation to check a repository install

For developers who work in a terminal, keep the skill beside the repository workflow it supports. Create a named folder under the project’s .claude/skills directory, with a SKILL.md file and any resources it needs. Open Claude Code in that project and invoke the skill by its slash name. Use a fresh project session to check what another developer would receive from the repository setup.

Ask for the same task in ordinary language after the explicit check passes. If the slash invocation works and the ordinary request misses the workflow, refine the description around the language teammates actually use. If the slash invocation fails, confirm the working directory and the skill’s folder and filename before revising the prompt.

Use a synthetic input that can be shared between the desktop and headless trials. Keep machine-specific absolute paths out of the comparison.

Extend Claude with skills - Claude Code Docs

Prove the headless job with fixed inputs

For an application or CI job, register the custom skill for API use and record its returned ID. Send a Messages API request with the skill’s type and ID in the container and include the code execution tool. Test through the same application path that will run in CI so you exercise its request construction and result handling.

In an illustrative trial, pass a CSV with three rows whose quantities are 4, 6, and blank. Define the expected result as a subtotal of 10 and a warning identifying row three. Run the fixture with the files and dependencies you intend to provide in production. A missing warning fails the output check. An unavailable dependency calls for verifying the target environment or preparing the input in your application.

Using Agent Skills with the API - Claude Platform Docs

Sources