Specs as contracts, not prompts: a full SDLC harness for your agent of choice

What it does

cc-sdd is a spec-driven development harness that installs 17 Agent Skills into your coding agent — Claude Code, Codex, Cursor, Copilot, Windsurf, OpenCode, Gemini CLI, or Antigravity. The flow is familiar: discovery, requirements, design, tasks, then implementation. The twist is the last step: kiro-impl runs each task with a fresh implementer doing TDD behind a feature flag, an independent reviewer, and an auto-debug pass that kicks in when the implementer stalls or the reviewer rejects twice. One task per iteration, and it’s safe to re-run after interruption.

The interesting bit

The philosophy is unusual: the spec is a contract between parts of the system, not a master command document. Code stays the source of truth; specs just make boundaries explicit so humans and agents can work in parallel without constant sync. Tasks carry Boundary: and Depends: annotations, and review checks for boundary violations, not just style nits. Learnings from earlier tasks propagate forward via Implementation Notes in tasks.md — a crude but effective memory mechanism.

Key highlights

• Same 17-skill set across all 8 supported agents; skills load on demand (progressive disclosure), no external dependencies
• kiro-discovery routes work: extend an existing spec, skip specs entirely, create one, or decompose a roadmap into multiple specs with cross-spec review
• EARS-format requirements, Mermaid-diagrammed design docs, and dependency-annotated tasks — typical spec output in under 10 minutes
• Templates and rules live in settings/ and are editable for PRD-style requirements, JIRA integration, or domain standards
• Existing Kiro specs remain compatible and portable; 13 languages supported

Caveats

• Only Claude Code and Codex integrations are stable — the other six are beta, meaning less real-world usage of subagent spawning and skill loading, with edge cases still surfacing
• Legacy /kiro:* command modes are deprecated, and the legacy Codex mode is outright blocked, so v1/v2 users need the migration guide
• The “under 10 minutes” spec-output claim is the README’s own, not a benchmark you can verify

Verdict

Worth a look if you’re doing feature-sized work with an agent and want structure heavier than vibes but lighter than a full process bureaucracy. If your changes are small and direct, discovery will just tell you to skip the spec — which is honestly the right answer.