Harness Copilot Skill: A GitHub Copilot Extension for Spec-Driven Development and Multi-Agent Workflows

Harness Copilot Skill project expands GitHub Copilot's capabilities to the field of spec-driven development (SDD). It uses a multi-agent workflow to intelligently convert requirement documents into code implementations, offering a new paradigm for AI-assisted development in software engineering teams. Key focus: moving AI from code completion to earlier stages of the software lifecycle (requirements & specs).

Traditional development follows linear "req→design→code→test" but often suffers from vague requirements, missing design records, leading to code-vs-intent deviations. Spec-driven development (SDD) addresses this by defining precise, verifiable system behavior before coding—specs as the single source of truth. Harness Copilot Skill bridges natural language requirements and formal specs, then auto-converts specs to code.

Harness uses a multi-agent architecture for different dev stages:

• Requirements Analysis Agent: Parses natural language requirements, identifies key concepts/constraints, asks clarifying questions, detects inconsistencies (e.g., for "thread-safe cache", queries size limits, eviction policies).
• Spec Generation Agent: Creates structured specs (Markdown/TLA+/Coq) with traceability links between requirements and spec elements.
• Code Synthesis Agent: Generates code based on specs (not context guesses) for more reliable results.
• Verification Agent: Checks code against specs via test cases, static analysis, formal verifiers; forms a "generate-verify-correct" loop.

Delivered as a GitHub Copilot extension (integrated into VS Code). Supports multiple spec formats (Markdown/TLA+/Coq) and languages (Solidity, TypeScript etc.). Key use cases:

• Smart Contracts: Ensures correctness (critical for blockchain apps).
• High-reliability systems: Meets regulatory requirements (aerospace, medical).
• Legacy modernization: Reverse-engineers specs from old code for refactoring.
• Team collaboration: Reduces communication gaps via shared specs.

• vs GitHub Copilot: Harness is "intent-driven" (based on specs) vs Copilot's "context-driven" (guesses from code), reducing understanding deviations.
• vs Cursor/Devin: Harness emphasizes reusable, verifiable specs (assets for long-term maintenance/reuse).

Current challenges:

• Learning curve: Developers need to learn spec writing.
• Spec maintenance: Keeping specs in sync with evolving code.
• Tool ecosystem: Formal verification tools lack generality/ease of use. Future plans: Improve spec assistance tools, integrate with version control, enhance error reporting to lower SDD barriers.

Harness Copilot Skill represents a key evolution—from "help write code" to "help define what correct code should do". This shift reflects a move to higher abstraction in software engineering. As AI advances, SDD may transition from academic concept to industrial standard for high-quality development.