spec-workflow

Provide a PRD issue, local PRD file, existing plan, or rough intent, then confirm a kebab-case slug. The workflow drafts spec.md, design.md, and tasks.md, dispatches parallel implementer subagents per group, and appends reviewer iterations to review.md until PASS.

Use it for non-obvious features where the plan itself is hard, work splitting into at least three parallelizable task groups, or cross-cutting changes needing a written spec and explicit reviewer pass. Skip it for single-file edits, trivial fixes, or exploratory spikes.

Yes, the playbook and artifacts are runtime-agnostic. Claude Code dispatches via the Task tool with implementer and reviewer subagent types; Codex uses its multi-agent tools, falling back to sequential orchestrator execution with the same briefs and status vocabulary.

Each failed review appends a fix iteration to tasks.md and the loop returns to the build phase. After three review cycles the workflow escalates to the user instead of looping a fourth time, since repeated failure signals the work needs reframing.

All four artifacts live under specs/<slug>/ from the project root, where the slug is kebab-case and confirmed with the user first. The skill never writes outside that directory, and review.md is append-only across iterations.