Enhancing AI coding assistants through open Spec-driven development (Spec-driven development for AI coding assistants), adopting the Claude Code Skills approach, compatible with various CLI and IDE...
概览
When the code or documentation generated by an AI coding assistant is disappointing, it's unclear whether it's because: 1. The large model's capabilities are problematic. 2. The AI coding assistant is problematic: providing inappropriate context (including system prompts). 3. The prompts I wrote are problematic. The focus is on breaking through "2. The AI coding assistant is problematic", but because the built-in spec-driven development methods (context engineering, system prompts, etc.) of various CLI and IDE AI coding assistants are not public, they are basically black boxes, making it impossible to analyze or optimize them yourself. I looked for various spec-driven development solutions, such as the 6A workflow, PRPs, Spec Kit, OpenSpec, spec-workflow-mcp, BMad-Method, etc., to see which one is the most suitable, and then proceed to optimize it myself. I roughly used these solutions and found that they each have their pros and cons.
README
以下为原文 README 节选,完整内容请查看原文。
Open Spec-Driven Development Skills (open-skilled-sdd)
Enhancing AI coding assistants through open Spec-driven development (Spec-driven development for AI coding assistants), adopting the Claude Code Skills approach, compatible with various CLI and IDE AI coding assistants that support AGENTS.md.
Why?
When the code or documentation generated by an AI coding assistant is disappointing, it’s unclear whether it’s because:
- The large model’s capabilities are problematic.
- The AI coding assistant is problematic: providing inappropriate context (including system prompts).
- The prompts I wrote are problematic.
The focus is on breaking through “2. The AI coding assistant is problematic”, but because the built-in spec-driven development methods (context engineering, system prompts, etc.) of various CLI and IDE AI coding assistants are not public, they are basically black boxes, making it impossible to analyze or optimize them yourself. I looked for various open spec-driven development solutions, such as the 6A workflow, PRPs, Spec Kit, OpenSpec, spec-workflow-mcp, BMad-Method, etc., to see which one is the most suitable, and then proceed to optimize it myself. I roughly used these solutions and found that they each have their pros and cons. My first thought was to create a comprehensive solution, integrating the advantages of each solution and removing the parts I thought were unsuitable. I soon discovered that this path was a dead end; this idea is exactly the channel that various AI coding assistants are currently working on. I cannot do better than these vendors. Moreover, from a philosophical point of view, a comprehensive solution (limited capabilities) cannot adapt to all types of scenarios (infinite scenarios). At this time, Claude Code Skills came out. I found this approach good; wouldn’t infinite skills be able to adapt to infinite scenarios? At the same time, I found skilled-spec.
My idea is to convert OpenSpec, PRPs, Spec Kit, spec-workflow-mcp, BMad-Method, 6A workflow, etc., including good spec-driven development methods that appear in the future, into individual composable Skills, which are open and can be optimized by yourself.
Skills
| Methodology | Skill | Trigger Words | Purpose |
|---|---|---|---|
| openspec | openspec-proposal-creation | “openspec proposal”, “openspec plan” | Generate structured proposals with EARS requirements |
推荐工具
换一个关键词,或者移除筛选条件。
安装
npx skillfish add forztf/open-skilled-sdd