A 660-Token AI Coding Workflow That Grills Your Requirements Before Writing a Line of Code

Most AI coding tools focus on generating code faster; this workflow attacks the real bottleneck—ambiguous requirements—by forcing structured interrogation before implementation. The 660-token footprint means it can run inside any LLM-powered coding agent without eating the context window needed for actual code.

The mattpocock/skills repository packages a four-step AI coding workflow—grill-with-docs, to-spec, to-tickets, and implement—that forces requirements clarity before any code is generated. Each skill is a short prompt that chains into the next: the AI interrogates the user one question at a time, resolves conflicts, writes Architecture Decision Records, and produces a structured spec with user stories, implementation decisions, and explicit out-of-scope boundaries. The entire skill set consumes only 660 tokens of context, making it one of the leanest AI coding frameworks available.

A new /wayfinder command extends the same interrogation pattern beyond code, mapping any open-ended goal—from a 10K running target to a product feature—into a dependency graph of sub-tasks and a concrete plan. The workflow treats context as a limited resource: after requirements are locked into markdown files, a /clear command wipes the chat history before /implement runs, preventing context pollution during code generation. /implement itself bundles test-driven development and code review, so the AI writes and critiques its own output in one pass.

The skills have been downloaded nearly 10 million times from skills.sh as of mid-July 2025, and the v1.1 release in early July tightened the combo sequence based on real usage. The core insight is that AI coding fails most often from underspecified requirements, not from weak code generation; a three-sentence grill prompt that asks one question at a time and records every decision produces more reliable output than a long, detailed system prompt.