Jev skills: install the instructions, test the decision

The official TypeSafe skill is a starting point for building Jev integrations. Installing it adds instructions, not API access or a replacement for your coding model. This independent Jagent guide separates installation, credentials and verification so you can tell whether a workflow is actually working.

Installation commands were checked against the official repository on 2026-11-03. Check its current requirements before upgrading an existing agent environment.

Install the official Jev skill

For Claude Code, the repository documents a marketplace plugin. Run these commands in the terminal where Claude Code is installed, then follow the host’s installation prompts:

claude plugin marketplace add typesafe-ai/skills
claude plugin install typesafe@typesafe-ai

For other supported coding agents, use the skills installer and select your agent when prompted. The repository describes project-local installation as the default; review the selected destination before choosing a global install.

npx skills add typesafe-ai/skills --skill typesafe-ai

Read the official SKILL.md to understand the instructions being installed. In Claude Code, the documented explicit invocation is /typesafe:typesafe-ai. For other hosts, confirm that the skill appears in that host’s loaded skills before asking it to implement an integration.

Give the skill a bounded first task

A useful request names the input, the allowed answers and the fallback. “Build an autonomous agent” leaves too many decisions unspecified. Start with one routing boundary whose outcome you can inspect.

Use the TypeSafe skill to build a support-ticket router.
Input: the ticket text.
Allowed queues: billing, technical, general.
Return the selected queue and the model's probabilities.
Keep queue execution in application code.
If the response is invalid, do not route automatically.
Show the request, response and a test for the fallback.

Review the generated implementation before connecting it to a live queue. A model selecting “billing” should not itself issue a refund. Keep permissions, validation and irreversible actions in the application’s control flow. Use examples where the expected queue is known, plus ambiguous tickets that should receive human review.

Try the judgment the skill would build around

This example uses Jagent’s existing live demo endpoint. It does not install a skill or run your local coding agent. Edit the ticket and run it to inspect a real routing judgment; normal demo quotas apply. No recorded result is presented as a fresh call.

Jev skill, MCP server or API?

You can use a skill to write API code without an MCP server. You can also connect an MCP server without installing a skill, provided the host knows how to use its tools. For Jagent’s hosted connection, follow the Jev MCP setup guide. Jagent credentials and TypeSafe credentials belong to different services; do not assume that either key works on the other endpoint.

Verify installation and the first response

1. Confirm the host loaded the installed skill. A successful package download alone does not show that the current session is using it.
2. Ask the agent to describe the bounded decision before generating code. Check that the available answers match your application.
3. Configure credentials for the selected provider outside tracked source files. The skill itself is not a credential.
4. Run one small request. Inspect the actual returned fields and record failures as failures.
5. Test an unavailable provider and malformed response. The application should retain a predictable fallback.

When a Jev skill does not work

If the host cannot find the skill, check installation scope and the host’s reload instructions. If the skill is loaded but the request returns an authorization error, check the provider, endpoint and key together. Reinstalling instructions will not fix an invalid credential. If tool discovery fails, investigate the MCP connection rather than the skill installer.

If a response picks an unexpected option, inspect the task description and option definitions. Overlapping choices make evaluation difficult. Build a small labeled set of real examples and review mistakes before selecting a confidence threshold; the example above does not establish a production accuracy rate.

Choose the next workflow

After a routing boundary works, reuse the same discipline for tool selection, relevance checks or context retention. Give each decision a defined input and output contract. A skill catalog helps discover workflows; it does not establish that every community project is compatible with your host or has been independently tested.