jevKnow what the skill changes
The TypeSafe agent skill helps a coding agent design and implement software that calls Jev. It gives the agent current guidance on the API, Choice, Score and Noul, workflow patterns and evaluation practices. It does not turn Jev into the language model behind Codex, Claude Code or another coding assistant. Jev still returns bounded, typed decisions; the coding agent still reads instructions, writes code and edits files.
That distinction matters when you choose a task. Ask the agent to add a routing, scoring, verification or review decision inside your application. Do not ask it to replace the conversational model with jev-latest.
Choose one installation method
For Claude Code, add the official typesafe-ai marketplace and install its TypeSafe plugin. For Codex and other supported agents, use the skills.sh installer and select the agent you actually run. The official instructions say to use one method so you do not create duplicate skill copies.
Installation alone does not make an API call and does not require your TypeSafe key. It adds guidance that the agent can read when it works on the project. Keep the installed copy current because an old skill can describe stale request or response fields.
# Claude Code claude plugin marketplace add typesafe-ai/skills claude plugin install typesafe@typesafe-ai # Codex or another supported agent npx skills add typesafe-ai/skills --skill typesafe-ai
Name the skill and the product behavior
Tell the coding agent to use the TypeSafe skill, then describe the behavior your product needs. A useful request names the input, the allowed outcome, the fallback and the evidence you will inspect. For example: route incoming support tickets to one of four teams, but keep uncertain or unmatched cases in human review.
The agent should work backward from that behavior. Exact rules, lookups, permissions and execution remain in code. Jev handles the semantic judgment. If several questions share the same state and do not depend on one another, the agent can ask them in one request and let code consume the relevant answers.
Review questions and thresholds in one place
Treat generated integration code as a draft. Review the state fields sent to TypeSafe, every instruction and criterion, all thresholds, the no-match behavior and the service-failure path. Keep those constants together so another engineer can audit the actual policy without searching through the whole application.
A typed result guarantees the response shape, not that the judgment is correct or that an action is authorized. Test representative and ambiguous examples. Keep raw probabilities available in your evidence, and measure review volume and expensive mistakes before allowing automatic behavior.
Keep keys and live tests under your control
In a web application, store the TypeSafe API key only on the server. Do not put it in a public environment variable or paste it into source files, logs or fixture exports. Automated tests should use recorded or mocked responses unless you deliberately approve live credit use.
When the agent proposes unfamiliar fields or a new SDK method, check the current TypeSafe documentation and the installed SDK types. The live documentation index is the source of truth; the skill is a guide for finding and applying it.
Update the skill when the contract moves
For the Claude Code plugin, update the marketplace and plugin, then reload plugins. For a skills.sh installation, run its update command. If an agent keeps inventing fields, confirm which copy it loaded before changing application code to match the invention.
After an update, rerun a small offline test and one deliberately approved live request before shipping. Record the actual returned model revision and preserve the failure response separately from a valid decision.
Prompt example
Using the TypeSafe skill, add one support-ticket routing decision. Keep exact permissions and execution in code, include an explicit review outcome, and test the policy with sanitized fixtures before any live action.