typesafe-ai: the skill that teaches agents typed decisions with Jev

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

Adding a little AI judgment to an application used to mean prompt-and-parse: write a prompt, make the model emit JSON, parse it, validate it, retry on failure. Type errors, hallucination, and retry loops are the fixed cost of that road. typesafe-ai is TypeSafe’s official single skill (MIT, repo created 2026-08-24, 2.7k stars as of 2026-10-12) that teaches an agent to build the same features from Jev’s typed judgments instead — the model returns choices, yes-or-nos, and probabilities, and code owns the workflow.

The order it works in

The skill fixes a four-step order for building with TypeSafe. First, read the live docs: the SKILL.md states outright that the live documentation is the source of truth, and passes on the trick that Mintlify-hosted docs serve Markdown when you append .md to the path, along with a task-to-page table for the programming model, primitives, confidence, and the SDKs. Second, work backward from behavior: decide what the application will show, select, change, or hand off, then derive the judgments it needs; known rules, calculations, exact lookups, and execution stay in code. Third, pick the primitive by what the answer means: Choice for one of a defined set, Noul for whether a condition holds, Score for a degree along a described dimension. Fourth, compose and verify: ask independent questions together in one request — including useful speculative questions with their premises stated explicitly — and drive behavior with probabilities and thresholds.

What it constrains

• Code owns the workflow: the model supplies programmable common sense only where ordinary code needs semantic understanding; routing, argument filling, evidence selection, and verify-and-escalate patterns compose around the goal instead of reaching for a universal classifier.
• One narrow judgment per question: independent dimensions may be split without destroying the relationship being judged, and candidate coverage gets checked — the model cannot choose an option you did not offer.
• Probabilities are thresholds, not truth: a Noul near 0.5 means similar probability for yes and no, not medium intensity, and Choice/Score confidence summarizes distribution concentration, not overall workflow correctness or permission to act.
• Typed output guarantees the interface, not truth: the SKILL.md’s own line, paired with the instruction to validate performance in the target domain and to treat cookbook thresholds and demo results as examples to evaluate.
• Credentials stay server-side: in web apps, API credentials must live on the server — written into the skill as a hard rule.

Who it’s for

Developers adding classification, routing, ranking, extraction, or verification — semantic if-statements — to their own applications. Once an agent has read it, “have the LLM generate and then parse” steps get rewritten as “code supplies facts, the model picks”. One honest catch: it needs a TypeSafe API key, and Jev is still in early access, so pricing, limits, and behavior can all change; the skill itself is a signpost, the method lives in docs that keep updating, and it neither writes your code nor guarantees the outcome. For what the community has built on Jev, read our ten Jev use cases roundup; for the wider Claude Code skill ecosystem, see Claude Skills Library.