Experimenting with Jev from Typesafe.ai using claude and their skill

In this experiment, I test the newly released typesafe.ai model—a highly efficient, generic classification model designed for deterministic decision-making rather than open-ended generative text. Using Claude Code, I built a Next.js application that leverages the typesafe.ai SDK to scan local project directories and categorize files as "safe to delete," "uncertain," or "keep" to help free up storage. By refining the prompt policy, the model successfully identified disposable files like node_modules and .next build caches, all while processing 200,000 tokens for just $0.01. 💡 Key Highlights: • Extreme Cost-Efficiency: Ran extensive directory testing and application design for a total API cost of exactly 1 cent. • Deterministic Output: Explored the model's boolean (yes/no), scoring metrics, and multiple-choice modes. • Practical Orchestration: Demonstrated how lightweight models are often better suited for routing and system cleanup than heavy generative models. 🛠️ Tech Stack: • typesafe.ai Jev model • Claude Code • Next.js 🔗 Links: • typesafe.ai: https://typesafe.ai/ • GitHub Repo for this project: https://github.com/mahan-ym/cleaner 👇 Connect with me: • https://x.com/mahan_ymt •   / mahan-ym   • https://github.com/mahan-ym