Descriptions:
Michael Arnaldi, creator of the Effect TypeScript library and founder of Effectful, leads a live unscripted workshop at AI Engineer on using AI coding agents to build and maintain complex library-level TypeScript code — territory most practitioners assume is too nuanced for vibe-coding to work reliably. Arnaldi, who hasn’t written code by hand since late 2024, applies the same AI-assisted approach to Rust and TypeScript libraries with zero documentation, zero online best practices, and highly specialized type machinery.
The session surfaces several concrete techniques. Spec-driven development: before any implementation, Arnaldi has the model produce a markdown specification file that serves as the persistent plan for the task, then implements against it in a separate pass. Context window loop automation: a simple bash script tells the model to pick up a small scoped task, implement it, and exit — then runs in a loop, preventing context degradation from accumulating unrelated history. The insight that less is more: Arnaldi describes internal Effectful research where reducing available tools significantly improved agent performance, with an extreme case being a single `execute` tool that runs arbitrary TypeScript (including shell via TypeScript) — forcing the model to use AST-based transformations rather than direct file patching, with surprisingly strong results.
The workshop also covers how to give agents access to undocumented libraries by cloning the repo directly into context, a pattern Arnaldi calls the most underrated technique in the space. The live coding is chaotic but the underlying workflow principles are clearly articulated and transferable.
📺 Source: AI Engineer · Published May 07, 2026
🏷️ Format: Hands On Build
