The parallel runner was supposed to build nix-key overnight. Instead I spent the day fixing bugs in the runner, documenting its state machine for a Temporal.io migration, and adding missing pieces to spec-kit while 75 of 77 tasks completed around me.
I dumped a rambling, typo-filled paragraph about turning a phone into a YubiKey into spec-kit's interview process. Three hours later: 101 functional requirements, 76 implementation tasks across 14 phases, and a parallel runner that immediately started breaking while I watched.
I wrapped spec-kit's SDD workflow in a Claude Code skill that adds research-driven interviews, disk-based fix-validate loops, structured test output, and a parallel agent runner. Here's how the skill works and what the output looks like on a real project.