Grill Me and Wayfinder: planning AI coding workflows

Matt Pocock explains how he turns his development process into reusable Markdown skills for a coding assistant. Grill Me asks detailed questions before implementation so the user can settle scope and design decisions. The host describes using it to discuss authentication and rate limits for an API endpoint. Pocock recommends this approach for substantial work that would be expensive to undo, while small fixes can proceed directly to implementation and review.

For larger projects, Pocock describes turning the discussion into a specification and dividing implementation into tickets, with one ticket per session. Wayfinder extends planning across sessions through a shared map of decisions, research and prototypes. He attributes this structure to his experience with context overload; his estimate of a roughly 150,000-token "smart zone" is a working judgment, not an established limit for every model.

The local versus cloud discussion concerns where development tools and agents operate. Pocock says he uses Discord to work with a remote server, which stays available for scheduled tasks. The episode does not demonstrate how to run models locally or establish offline inference support.

He also describes using terms such as "tracer bullet" and "vertical slices" to guide an AI agent toward earlier feedback. On testing, he favors evidence that a test fails without the change and warns that generated tests can merely repeat the implementation.