specvibe（specvibe 开发）

SpecVibe: The AI-Native Development Framework

This skill provides a universal, seven-stage framework for developing production-ready, AI-native applications. It enforces a "Specification-as-Source-of-Truth" mindset, ensuring that every aspect of the project is defined, testable, secure, and documented before and during implementation, following the most advanced 2026 community best practices from Google, GitHub, and Thoughtworks.

Core Philosophy

• Intent is the Source of Truth: The specification (spec.md) is the primary artifact. Code is the last-mile implementation of that intent.
• Human-AI Collaboration: Follow the Delegate/Review/Own model at every stage to maximize efficiency and maintain quality.
• Iterate in Small, Validated Chunks: Break down work into the smallest possible units, test them, and commit frequently. Never let the AI generate large, monolithic blocks of code.
• Automate Everything: Use tests, linters, CI/CD, and automated documentation to build a robust quality assurance system.

The Seven Stages of AI-Native Development

Follow these stages sequentially. Each stage has a Quality Gate—a set of questions you must answer before proceeding—and a clear Delegate/Review/Own model for human-AI collaboration.

Stage 1: Specify - The Intent

Goal: Define *what* to build and *why* in a structured spec.md.

• Delegate: Ask the AI to interview you about the project goals and generate a draft spec.md using the templates/spec-template.md.
• Review: Check if the spec accurately captures all user stories, edge cases, and success metrics.
• Own: The final approval of the user requirements and business goals.

Quality Gate 1: Specification Review

• *Does the spec.md clearly define the user, their problem, and the proposed solution?*
• *Are non-functional requirements (performance, security, accessibility) listed?*
• *Is the scope well-defined and unambiguous for an AI to understand?*

Stage 2: Plan - The Blueprint

Goal: Translate the spec.md into a concrete technical plan.

• Delegate: Feed spec.md to the AI and ask it to generate a PLAN.md detailing the architecture, data models (using references/01-schema-and-types.md), API contracts (using templates/openapi-template.yaml), and a task breakdown.
• Review: Assess the proposed tech stack, architecture, and task list for feasibility and alignment with best practices.
• Own: The final architectural decisions and technology choices.

Quality Gate 2: Plan Review

• *Is the chosen architecture appropriate for the project's scale and requirements?*
• *Is the API contract complete and consistent with the data models?*
• *Are the tasks small, independent, and logically sequenced?*

Stage 3: Test - The Safety Net

Goal: Define the application's behavior through a comprehensive, failing test suite.

• Delegate: Ask the AI to generate a full suite of tests (unit, integration, E2E) based on spec.md and PLAN.md. Refer to references/05-testing.md.
• Review: Ensure tests cover all user stories, API endpoints, and critical business logic. Check for meaningful assertions.
• Own: The definition of "done" for each feature, as represented by the tests.

Quality Gate 3: Test Suite Review

• *Does every feature in the spec have corresponding tests?*
• *Do all tests currently fail for the correct reasons?*

Stage 4: Implement - The Engine Room

Goal: Write clean, efficient code that makes all tests pass.

• Delegate: Instruct the AI to implement one task at a time, feeding it the relevant spec, plan, and failing test. Use the "chunked iteration" strategy from references/08-ai-collaboration.md.
• Review: After each small chunk, review the generated code for correctness and style. Do not wait for the entire feature to be complete.
• Own: The responsibility for committing each validated chunk of code to version control.

Quality Gate 4: Implementation Review

• *Do all tests for the implemented task now pass?*
• *Is the code clean, readable, and consistent with the project's style guide?*
• *Has the change been committed to Git with a clear message?*

Stage 5: Review - The Quality Shield

Goal: Ensure the implemented code is secure, robust, and maintainable.

• Delegate: Automate security scans (SAST, DAST, dependency checking) in CI. Use an AI agent to perform a preliminary code review based on references/04-security.md (OWASP 2025).
• Review: A human developer must perform a final review, focusing on logic, architecture, and subtle bugs that AI might miss.
• Own: The final approval (LGTM) to merge the code into the main branch.

Quality Gate 5: Code Review

• *Does the code pass all automated security and quality checks?*
• *Has a human engineer reviewed and approved the changes?*

Stage 6: Document - The Knowledge Base

Goal: Create clear, comprehensive documentation for both users and developers.

• Delegate: Use AI to generate initial drafts of API documentation from the OpenAPI spec, and user guides from the spec.md. Refer to references/09-documentation.md.
• Review: Edit the AI-generated content for clarity, accuracy, and tone. Add diagrams and examples.
• Own: The final, published documentation that serves as the official source of information.

Quality Gate 6: Documentation Review

• *Is the API documentation accurate and complete?*
• *Is the user guide easy for a non-technical person to understand?*

Stage 7: Deploy - The Launchpad

Goal: Automate deployment and ensure the application is observable and reliable in production.

• Delegate: Ask the AI to generate Dockerfiles, CI/CD pipeline configurations (e.g., GitHub Actions), and infrastructure-as-code scripts. Refer to references/06-devops.md.
• Review: Verify the deployment scripts, container configurations, and monitoring setup (references/07-error-handling.md).
• Own: The production environment and the ultimate responsibility for uptime and reliability.

Quality Gate 7: Production Readiness Review

• *Can the application be deployed and rolled back with a single command?*
• *Is comprehensive, structured logging (OpenTelemetry) in place?*
• *Are alerting and monitoring configured for key performance indicators?*