when agents do the typing, the bottleneck moves. the scarce thing is no longer keystrokes — it is a clear, reviewable description of what should be true. so we write that first, and let the code follow.
the pipeline
the whole studio runs on claude-plugin-sdd. a decision becomes an ADR. an ADR becomes an OpenSpec spec with RFC-2119 requirements. the spec gets planned into tracked issues, and a team of agents implements them in parallel git worktrees — then reviews and merges its own PRs.
/sdd:adr → decide
/sdd:spec → specify
/sdd:plan → break down
/sdd:work → build (parallel worktrees)
/sdd:review → merge
what changes
code review becomes spec review. a regression is usually a spec that was silent about a case, not a fumbled keystroke. and because every artifact is versioned, you can trace any line back through markup → brief → dispatch → commit.
a regression is usually a spec that was silent, not a fumble.
it is slower to start and much faster to sustain. the spec is the thing you actually maintain.