OpenSpec tutorial: plan coding tasks with Pi and MiniMax M3

OpenSpec turns a product requirements document or an existing codebase into specifications and actionable coding tasks. This tutorial follows a new project through installation, agent setup, exploration, and a proposal for an initial MVP. The speaker describes it as an open source planning framework whose documents stay in the repository across sessions and agents.

The setup starts with an npm installation and OpenSpec init inside the project. The speaker selects Pi, which installs the project’s skills, then connects MiniMax M3 using an API key. A YAML configuration accepts project context and rules, though the demonstration leaves it alone. This AI agent workflow uses an external model provider; it does not demonstrate running a model locally.

Starting from a detailed PRD, Pi recommends exploration before a proposal. The resulting MVP change contains a design, a proposal, and almost 50 tasks with review checkpoints. OpenSpec also has commands to inspect active work, apply tasks, merge delta specs, and archive completed changes.

The example shows why the generated plan still needs engineering review. Without explicit stack requirements, the agent chooses Next.js, Drizzle, and Postgres, which the speaker says would not be their first choice. The speaker reports favorable token use and cost with MiniMax M3, a sponsored provider in this video, but the demonstration is not a controlled model comparison.