devenjarvis/lathe

What it solves

Lathe is designed to prevent LLMs from simply doing the work for the user, instead using them to create personalized, hands-on technical tutorials. It solves the problem of "zero-to-one" learning in obscure or new domains where human-written tutorials are scarce, allowing users to internalize concepts by manually typing and working through the generated guides.

How it works

Lathe combines a Golang CLI and a set of LLM "skills" (prompts/commands) that integrate with existing coding agents (like Claude Code, Cursor, or Windsurf).

• Generation: The user prompts a coding agent using a /lathe skill to generate a multi-part technical tutorial.
• Management: The Go binary manages the storage of these tutorials in ~/.lathe/tutorials/ and serves them via a local web UI at http://localhost:4242.
• Interaction: The UI allows users to ask questions, verify if the tutorial works, or extend it with new parts. These actions are driven by the coding agent via a worker loop (/lathe-work) or manual copy-paste commands.
• Customization: Users can define "voices" to control the prose style of the tutorials.

Who it’s for

Software engineers and learners who prefer hands-on learning and want to dive into new technical domains without relying on the LLM to write the code for them.

Highlights

• Agent-Agnostic: Works with a wide range of coding agents including Claude Code, Cursor, Codex, Gemini CLI, opencode, Cline, and Windsurf.
• Local-First UI: A purpose-built reader with table of contents navigation, side-notes for deeper thinking, and exercises at the end of each part.
• Provenance Tracking: Every tutorial records the specific LLM model, the "voice" used, and the research sources (URLs) consulted during generation.
• Interactive Learning: Includes built-in functionality to verify tutorials, ask questions about specific parts, and add new parts to a series.