Use case · spec-first
Ship Features Spec-Driven with AI
The disciplined alternative to vibe-coding: write a clear spec, let the AI agent implement to it, verify with tests, ship. Setup in 5 minutes using skills-hub's spec-driven composition. Works in Claude Code, AWS Kiro, Cursor, Antigravity 2.0.
Short answer
Install the `story-implementer` composition. Write a clear spec (requirements + acceptance criteria + edge cases). Invoke /story-implementer with the spec file. The agent implements to spec, runs tests against it, opens a PR. Halts on any unmet acceptance criterion. AWS Kiro makes this native; Claude Code + Cursor + Antigravity do it via the skill.
Spec-driven shipping is the disciplined alternative to vibe-coding. You write the spec; the agent implements to it; tests verify it. The discipline shows up in three places: clearer acceptance, fewer rewrites, and an actual paper trail for what 'done' means.
Step-by-step
- 1
Install the spec-driven composition
story-implementer chains spec → implement → test → review → PR. Installing it pulls every child skill automatically.
$ npx @skills-hub-ai/cli install story-implementer - 2
Write the spec
Use the /spec skill to generate a structured spec from a feature description. Or write one by hand, requirements, acceptance criteria, edge cases, non-functional requirements.
$ npx @skills-hub-ai/cli install spec - 3
Invoke the implementation pipeline
Run /story-implementer pointing at the spec file. The agent team takes over: implement → test → review → PR. Each stage is an isolated subagent (on Claude Code) or a Cascade flow step (on Windsurf).
- 4
Review the PR
The agent surfaces a single PR with the diff, tests, and a markdown summary of which acceptance criteria are met. Either merge or send back with specific feedback.
- 5
Iterate if criteria are unmet
If any acceptance criterion isn't met, the agent halts and asks. You can extend the spec or send specific feedback; the agent re-runs only the failing portion, not the whole pipeline.
Skills installed in this recipe
Frequently asked questions
Why spec-driven vs vibe-coding?
Vibe-coding wins for exploration and prototypes. Spec-driven wins for production paths where 'done' has a clear definition and rewrites are expensive. Most teams use both, spec-driven for revenue-critical features, vibe for spikes.
What's a good spec format?
Requirements (what the feature does), Acceptance criteria (testable conditions), Edge cases (failures, empty states, race conditions), Non-functional requirements (perf, accessibility, security), and an explicit Out-of-scope section.
Which AI tool is best for spec-driven?
AWS Kiro makes it native to the IDE. Claude Code via /story-implementer composition is the strongest non-Kiro option. Cursor + Antigravity 2.0 work via the same skill. See /compare/kiro-vs-cursor for the trade-offs.
How long does spec-driven take?
Spec writing: 20-90 min for a typical feature. Agent implementation: 5-30 min for small features, 1-3 hours for large. Total wall-clock often beats vibe-coding because there are fewer rewrites.
What if the spec is wrong?
The agent halts at the first unmet acceptance criterion. That's the spec being wrong, not the agent. Update the spec, re-run the failing portion. The cost is much lower than discovering the spec was wrong after merging vibe-coded code.
Related use cases
Automate Pull Request Review with AI
Install `code-review`, `secure`, `unit-test` skills via `npx @skills-hub-ai/cli`. Add a GitHub Actions workflow that run…
Migrate a Legacy Codebase with AI Agents
Install the `migrate-safe` composition. Run it against your repo with the target framework/language as the argument. The…
Browse 4,900+ skills in the catalog
Free, MIT, works in every major AI coding tool.
Browse the catalog →