Jev With Claude Code and Codex: The TypeSafe Agent Skill
Last checked · Independent guide, not affiliated with TypeSafe AI
TypeSafe publishes an official agent skill that teaches coding agents how to write Jev calls. In Claude Code, install it with 'claude plugin marketplace add typesafe-ai/skills' and 'claude plugin install typesafe@typesafe-ai'; for Codex and other agents, run 'npx skills add typesafe-ai/skills --skill typesafe-ai'. Then ask the agent to use the TypeSafe skill.
Jev does not write code itself, but coding agents are good at writing code that calls Jev. TypeSafe ships an official agent skill so that agents such as Claude Code and Codex know the API, the three question types, the recommended patterns and TypeSafe’s best practices without you pasting documentation into the chat.
What the skill does
Section titled “What the skill does”The skill is a folder of instructions that your coding agent loads when it works on Jev-related code. According to TypeSafe, it covers:
- the request format and the Noul, Choice and Score question types;
- architectural patterns such as asking many questions per call and gating on confidence;
- how to structure evaluations so questions and thresholds are easy to review.
TypeSafe’s docs make one point about why this matters: coding agents tend to fall into a one-question-per-call habit, which is slower and costs more. The skill tells them to pack questions into each request instead.
Install in Claude Code
Section titled “Install in Claude Code”Run these in your terminal:
claude plugin marketplace add typesafe-ai/skillsclaude plugin install typesafe@typesafe-aiYou can then invoke it directly with /typesafe:typesafe-ai, or just mention “the TypeSafe skill” in a prompt.
To update later:
claude plugin marketplace update typesafe-aiclaude plugin update typesafe@typesafe-aiRestart Claude Code or run /reload-plugins to load the new version. You can also turn on automatic updates for the marketplace from the /plugin menu.
Install in Codex and other agents
Section titled “Install in Codex and other agents”npx skills add typesafe-ai/skills --skill typesafe-aiThe installer asks which agent to install for. Installation is per project by default; add -g to install globally. Update with npx skills update. You can also copy the skills/typesafe-ai folder from the GitHub repository by hand. Pick one method only, so you do not end up with two copies.
Prompts that work
Section titled “Prompts that work”Name the skill in the prompt; that works in any agent. Three starting points adapted from TypeSafe’s suggestions:
- Find opportunities: “Using the TypeSafe skill, look through this project for places where a quick judgment could replace fragile parsing or long if-else chains, and list them with the question type you would use.”
- Experiment cheaply: “Using the TypeSafe skill, run a few test queries with the key in
TYPESAFE_API_KEYand propose changes based on what works.” At Jev’s prices, dozens of test calls cost a fraction of a cent. - Apply a known pattern: point the agent at a specific TypeSafe cookbook, such as citation checking or re-ranking, and ask whether it fits your code.
Keep the agent honest
Section titled “Keep the agent honest”TypeSafe’s own advice for working with agents on Jev code:
- Review the plan before the code. Agents are not good at writing questions; expect to edit the question text yourself.
- Keep questions and thresholds in one file. That is the part a human needs to review, and the part you will tune.
- Do not accept claims untested. Ask the agent to validate assumptions with real calls.
- If the agent invents request or response fields, the skill is probably out of date. Update it and try again.
Common problems
Section titled “Common problems”| Problem | Fix |
|---|---|
| The agent ignores the skill | In Claude Code, invoke /typesafe:typesafe-ai; elsewhere, say “use the TypeSafe skill”. Check the installer targeted the agent you are using, then restart it |
| Routing gives too many false positives or negatives | Thresholds are too low or too high, or the questions are vague. Tighten the question before moving the threshold |
| Confidence thresholds everywhere | If you just need the best option, take choice directly; use probabilities for statistical work |
| Requests fail with 400 “Unknown model” | Make sure the generated code uses jev-1.13.0 or jev-latest, not jev-1.13. See 400 |
Jev inside the coding loop
Section titled “Jev inside the coding loop”Separately from writing Jev code, developers have started using Jev as part of coding workflows: routing tasks to the right coding model, running yes/no review checks on diffs, and deciding when an agent’s step is finished. A YouTube demo pairing Jev with Claude Code as a cheap decision layer drew tens of thousands of views in the first week. The same limits apply there as anywhere: Jev decides, and the LLM writes. See What can Jev do?