Deep Analysis

针对 AI 加速开发的文档优先方法论，通过结构化的 4 个阶段将 AI 生成代码从"运气"变为"确定性"，实现 10-20 倍生产力提升

Core Features

Technical Implementation

Highlights

解决"速度幻觉"问题：AI 工具提高打字速度但不能加快项目交付
实际案例验证：5Levels 项目 46 个智能端点在 4.5 小时内完成
有学术研究支撑：引用 McKinsey、DORA、METR 等机构的 2025 年研究数据
核心哲学创新："代码失败时，修复规范而非代码"

Use Cases

技术创始人快速构建 greenfield 产品
1-5 人小团队提高 AI 辅助开发效率
后端和业务逻辑密集的项目开发
规范驱动开发(SDD)的实践落地

Limitations

不适合大型企业
前端开发场景需要补充 AI 设计工具配合使用
要求团队有较强的文档意识和执行纪律
不适合黑客松或一次性原型项目

Tech Stack

Claude/Claude CodeCursor IDEWindsurfGitHub CopilotMarkdown规范驱动开发(SDD)框架

README

View on GitHub

Stream Coding v3.4

Name: stream-coding
Rating: 3.7 (3 reviews)
Author: frmoretto

The 10-20x Methodology for AI-Accelerated Software Development

📢 v3.4 Update: Complete 13-item Clarity Gate (was 5 items), scoring rubric, Documentation Audit. Re-download if using older version.

"Stream coding isn't about faster coding. It's about documentation so clear that code writes itself."

The Problem: The Velocity Mirage

AI tools promise 10x productivity. GitHub Copilot, Cursor, Claude Code—they make coding 55% faster.

But projects still take the same time to ship.

Why? Because faster typing doesn't solve:

Strategic decisions AI can't make for you
Context that gets lost between prompts
Technical debt created at 10x speed

This gap between task velocity and project velocity is the Velocity Mirage.

The Solution: Stream Coding

Stream Coding is a documentation-first methodology that makes AI-generated code deterministic.

The 40/40/20 Split:

40% Strategic Thinking (Phase 1) — Solve hard problems before coding
40% AI-Ready Documentation (Phase 2) — Specs so complete AI has zero decisions
20% Execution + Quality (Phases 3-4) — Code streams out automatically

Real Results (5Levels Case Study):

7 production modules in 4.5 hours
46 intelligence endpoints (77 total backend API)
Zero bugs in generated code, 21 minutes average per tested module

Note: The case study focuses on backend intelligence modules—Stream Coding's sweet spot. For frontend, use the methodology for behavior (components, state, logic) and complement with AI design tools for visuals. See Chapter 4 for details.

Quick Start

Option 1: claude.ai / Claude Desktop

Download stream-coding.zip
Go to Settings → Features → Skills → Add
Upload the zip file
Ask Claude: "Build a user authentication system"

Option 2: Claude Code

Copy SKILL.md to your project's skills folder
Claude Code will automatically detect and use it
Ask Claude: "Build a user authentication system"

Option 3: Claude Projects

Add SKILL.md to project knowledge. Claude will search it when needed, though Skills provide better integration.

Option 4: Manual / Other LLMs

For Cursor, Windsurf, or other AI tools:

Extract core principles (Phases, Document Types, Clarity Gate)
Create a condensed version for .cursorrules or project settings
Use the templates and Clarity Gate Checklist as reference

The methodology is tool-agnostic—only SKILL.md is Claude-optimized.

Read the Manifesto

The /manifesto folder contains the complete methodology:

Chapter	Topic
Chapter 1	The Velocity Mirage
Chapter 2	Why AI Tools Alone Fail
Chapter 3	The Missing Middle
Chapter 4	The 4-Phase Methodology
Chapter 5	Day 2 & The Rule of Divergence
Appendix A	Templates & Checklists
Appendix B	Research & SDD Positioning
Appendix C	5Levels Case Study (Git-Verified)
Advanced Framework	Document Architecture (v3.3)

Use the Templates

The /templates folder contains ready-to-use frameworks:

Template	Purpose
Strategic Blueprint	Answer the 7 Phase 1 Questions
ADR Template	Document architecture decisions with rationale
Clarity Gate Checklist	The mandatory Phase 2→3 gate

The Core Insight

"When code fails, fix the spec—not the code."

Traditional development iterates on code. Stream Coding iterates on documentation.

Every manual code edit without updating the spec creates Divergence—technical debt that breaks the stream. The methodology works because it treats code as a compiled output of documentation, not the source of truth.

Who This Is For

✅ Technical founders building greenfield products
✅ Solo developers and small teams (1-5 people)
✅ Anyone tired of AI-generated spaghetti code
✅ Backend/business logic focused (see Chapter 4 for frontend approach)

❌ Not for large enterprises (see GitHub Spec-Kit, Kiro, Gemini Conductor)
❌ Not for hackathons or throwaway prototypes
❌ Not for teams who can't commit to documentation-first

Research Validation

Stream Coding aligns with industry research:

McKinsey (2025): Top performers see 16-30% productivity gains through "end-to-end PDLC implementation" and "structured communication of specs"
DORA (2025): 7.2% delivery instability increase for every 25% AI adoption without foundational systems
METR (2025): Developers 19% slower with AI despite feeling 20% faster

The methodology isn't magic, it's systematic application of spec-driven development at founder scale.

Contributing

Stream Coding is open source under CC BY 4.0. You're free to use, adapt, and share with attribution.

Found an improvement? Open an issue or PR.

About

Created by Francesco Marinoni Moretto while building 5Levels, a LinkedIn relationship intelligence platform.

The methodology emerged from building 7 production modules in 4.5 hours, and documenting exactly how.

License

CC BY 4.0 — Use freely with attribution.

"Stream Coding methodology by Francesco Marinoni Moretto (github.com/frmoretto/stream-coding)"

In memory of my beloved father Guido