vibe-coding-essentials

人工智能辅助开发的工作流框架。四种操作模式，25条黄金法则，以及针对9个框架的防幻觉护栏。

  • 默认分支main
  • 大小78 KB
  • 更新7月6日
在 GitHub 打开

README

Vibe Coding Essentials

Version License

🏆Recognized: Awesome Vibe Coding: Officially featured in the premier directory of vibe coding resources!

Most developers use the same workflow for every AI task.

That is the mistake.

Different work requires drastically different AI behavior. Wrong mode = wrong outcomes.

Has this happened to you?

□ AI created 12 files for a button component

□ AI added 4 packages to format a date

□ Context window exploded mid-feature

□ API bill was ridiculous for a session that shipped nothing

□ Day 2 AI forgot everything decided on Day 1

□ The code works but nobody understands it

The Four Operating Modes

Mode Goal Accepts Refuses
⚡ Economy Minimize cost Simpler solutions Large context scans
🚀 Builder Ship fast Some technical debt Premature abstractions
🛠 Maintainer Preserve stability Incremental improvements Refactors & new dependencies
🏛 Architect Make decisions Planning & tradeoffs Production code generation

Why Modes Matter

Debugging in Builder Mode creates unreviewed refactors.

Designing in Maintainer Mode creates over-engineered fixes.

Investigating in Economy Mode creates context collapse.

Coding in Architect Mode creates premature implementations.

Using the wrong mode creates:

  • technical debt
  • token waste
  • regressions
  • unnecessary complexity

Mode Switching

Healthy AI-assisted development involves switching modes as work evolves.

🏛 Architect
      ↓  (Decision approved)
🚀 Builder
      ↓  (Feature complete)
🛠 Maintainer
         (Production ready)

Performance & Benchmarks

Strict behavioral rules don’t just prevent bugs—they reduce code bloat. In a recent benchmark against a qwen3:8b local model evaluating a standard coding task (Node.js filesystem ops):

  • Baseline (Raw Model): 31 Lines of Code
  • Ponytail Base Prompt: 21 Lines of Code
  • Ponytail + Vibe Coding Essentials: 17 Lines of Code

Injecting our anti-hallucination guardrails shaved an additional 19% off the code length compared to the base “lazy dev” prompt by forcing the model to adhere strictly to native APIs instead of inventing unnecessary abstractions.

What This Repository Contains

How to Use

1. Install the runtime Copy AGENTS.md to your project root. (Tool-specific versions are available in templates/).

2. Set your context Update the 3-line MVP Context block at the top of AGENTS.md daily.

3. Command the AI Builder is the default behavior. To change how the AI operates for a specific task, start your prompt with a mode tag:

“[ECONOMY] Why is the validateToken function throwing a TypeError?”

“[MAINTAINER] We have a regression in the billing calculator. Write a failing test first.”

“[ARCHITECT] Let’s design the database schema for the new auth flow.”

4. Utilize the Framework The files in this repository are designed for distinct parts of your workflow:

  • anti-hallucination/: Find your framework, copy the rules, and paste them into your AGENTS.md so the AI stops using deprecated APIs.
  • prompts/: When the AI starts stuck in a correction loop, stop. Open the relevant template here to hard-reset the context.
  • SKILL.md & GOLDEN_RULES.md: Read these once. Treat them as the Clean Code of the AI era. They teach the meta-skills of driving an agent.
  • LESSONS_LEARNED.md & production/hardening.md: Review these before shipping. They document the catastrophic AI failures you must manually check for.

Star History Chart

Philosophy

The model generates.

The developer decides.

You own the workflow.

评论

暂无评论