Vibe Coding with LLMs

A practical guide to building real software with LLMs while maintaining correctness, stability, and long-term maintainability.

Get Started View on GitHub

Core Philosophy

Vibe coding works because it replaces “trusting the model” with systems that enforce correctness. You move fast without breaking things—not because the LLM is perfect, but because your process is.

Treat the LLM as a junior engineer with infinite stamina: fast, tireless, and helpful—but lacking long-term judgment unless you supply it.

Your job is to provide:

  • Constraints - Define what’s allowed and what’s not
  • Invariants - Rules that must always hold
  • Tests - Automated verification of correctness
  • Clear expectations - Explicit requirements and conventions

Getting Started

New to Claude Code?

Start with the comprehensive setup guide:

Claude Code Setup Guide

Covers terminal setup, running multiple Claudes, Plan mode, CLAUDE.md, slash commands, hooks, and more.

Essential Reading

Guide What You’ll Learn
Quick Reference Checklist Print-friendly checklist for daily use
Incorporate into Your Repo Agent-facing checklist to adopt these practices in any repo
Anti-Patterns & Warning Signs When AI development goes wrong
Audit Findings & Lessons Real security audit of AI-generated code
Automation & Testing CI/CD, hooks, and verification loops
Context Management Managing LLM sessions effectively
Code Review for AI Reviewing AI-generated code
Setting Up AI Code Review Tutorial: zero to automated PR reviews
Version Control Git workflow for AI development
MCP Tool Grouping Scaling MCP servers beyond 20 tools
Plan-Driven Development Structured planning workflow for coding agents
Dependency Safety Evaluating and managing AI-suggested packages
Refactoring with AI Changing code safely with AI assistance
Smart Contract Auditing Multi-agent audit workflow for Solidity contracts
Auditing at Scale External tools for large AI-generated codebases

The Workflow

1. Start with a Product Conversation

Begin by talking to the LLM about what you want to build, not how yet. Describe the problem, users, and environment.

2. Co-Design the Prompt

Ask the LLM to help refine requirements before implementation. Identify missing constraints and trade-offs.

3. Produce a Roadmap

Generate phases with clear goals, features, and exclusions. Phases reduce scope creep.

4. Create Architecture Documents

Generate architecture and README docs early. These anchor consistency across sessions.

5. Define Invariants

Create non-negotiable rules that must always hold. Reference before any implementation.

6. Add Tests Immediately

Introduce tests as soon as real code appears. Tests are gates, not optional feedback.

7. Optimize for Boring Code

Tell the LLM explicitly: boring, predictable, production-grade solutions. Creativity is opt-in.

8. Favor Modularity

Keep files under ~1,500 lines. Smaller modules improve correctness and AI comprehension.

9. Run Bug Hunts

Regularly ask the LLM to review changes, check edge cases, find regressions, and audit data at component boundaries — where AI code most often breaks.

10. Give Claude Verification

The most important pattern: Give Claude a way to verify its work. This 2-3x the quality of output.

graph TD
    A([Define Constraints & Invariants]) --> B[Generate Code]
    B --> C{Automated <br/>Verification}
    C -- No --> D[Refactor based on Error]
    D --> B
    C -- Yes --> E([Merge & Maintain Velocity])

    style A fill:#e1f5fe,stroke:#01579b
    style C fill:#fff9c4,stroke:#fbc02d
    style E fill:#c8e6c9,stroke:#2e7d32

Templates

Prompts Library

Why This Works

Traditional software development emphasizes careful upfront planning because writing code is expensive. With LLMs, code generation is cheap—but maintaining correctness is still hard.

This workflow inverts the traditional approach:

  • Spend time on constraints, tests, and documentation
  • Let the LLM handle the tedious implementation
  • Use automated checks to catch drift

The result: velocity without chaos.

Contributing

This guide is a living document. If you’ve found patterns that work (or anti-patterns that don’t), please contribute:

  1. Share your templates in /docs/templates
  2. Add useful prompts to /prompts
  3. Document lessons learned

This guide is released into the public domain. Use it however helps you build better software.

Back to top

Released into the public domain. Use it however helps you build better software.