Ultimate Guide to Vibe Coding V1.2

Author: Nicolas Zullo, https://x.com/NicolasZu
Creation Date: March 12, 2025
Last Update Date: October 06, 2025

---

Getting Started

To begin vibe coding, you only need one of these two tools:

• Claude Sonnet 4.5, in Claude Code
• gpt-5-codex (high), in Codex CLI

This guide works for both the CLI versions (to use in the terminal) and the VSCode extension versions (both Codex and Claude Code have one, with a more recent interface).

(Note: While earlier versions of this guide utilized Grok 3, we then transitioned to Gemini 2.5 Pro. And now we're using Claude 4.5 (or gpt-5-codex (high)))

(Note 2: If you want to use Cursor, please check the version 1.1 of this guide, but we believe it's less powerful than Codex CLI or Claude Code)

Setting up everything correctly is key. If you’re serious about creating a fully functional and visually appealing game (or app), take the time to establish a solid foundation.

Key Principle: Planning is everything. Do NOT let the AI plan autonomously, or your codebase will become an unmanageable mess.

---

Setting Up Everything

1. Game Design Document

• Take your game idea and ask GPT-5 or Sonnet 4.5 to create a simple Game Design Document in Markdown format: game-design-document.md.
• Review and refine the document to ensure it aligns with your vision. It’s fine if it’s basic—the goal is to give your AI context about the game’s structure and intent. Do not over-engineer as we will iterate later.

2. Tech Stack and CLAUDE.md / Agents.md

• Ask GPT-5 or Sonnet 4.5 to recommend the best tech stack for your game (e.g., ThreeJS and WebSocket for a multiplayer 3D game). Save this as tech-stack.md.
  • Challenge it to propose the simplest yet most robust stack possible.
• In your terminal, open Claude Code or Codex CLI and use the /init command. It will use the two .md files you created so far. This will create a set of rules so your LLM is guided correctly.
• Crucially, review the generated rules. Ensure they emphasize modularity (multiple files) and discourage a monolith (one giant file). You might need to manually tweak or add rules. Review also when they trigger.
  • IMPORTANT: Some rules are critical for maintaining context and should be set as "Always" rules. This ensures the AI always refers to them before generating code. Consider adding rules like the following and marking them as "Always":

    # IMPORTANT:
    # Always read memory-bank/@architecture.md before writing any code. Include entire database schema.
    # Always read memory-bank/@game-design-document.md before writing any code.
    # After adding a major feature or completing a milestone, update memory-bank/@architecture.md.
    

  • Example: Ensure other (non-"Always") rules guide the AI towards best practices for your stack (like networking, state management, etc.).
  • This overall rules setup is mandatory if you want a game that is as optimized as possible, and code as clean as possible.

3. Implementation Plan

• Provide GPT-5 or Sonnet 4.5 with:
  • The Game Design Document (game-design-document.md)
  • The tech stack recommendations (tech-stack.md)
• Ask it to create a detailed Implementation Plan in Markdown (.md) which is a set of step-by-step instructions for your AI developers.
  • Steps should be small and specific.
  • Each step must include a test to validate correct implementation.
  • No code—just clear, concrete instructions.
  • Focus on the base game, not the full feature set (details come later).

4. Memory Bank

• Create a new folder for your project and then open it in VSCode.
• Inside the project folder, create a subfolder named memory-bank.
• Add the following files to memory-bank:
  • game-design-document.md
  • tech-stack.md
  • implementation-plan.md
  • progress.md (Create this empty file for tracking completed steps)
  • architecture.md (Create this empty file for documenting file purposes)

---

Vibe Coding the Base Game

Now the fun begins!

Making sure everything is clear

• Open Codex or Claude Code in VSCode's extensions or launch Claude Code or Codex CLI in the terminal of your project.
• Prompt: Read all the documents in /memory-bank, is implementation-plan.md clear? What are your questions to make it 100% clear for you?
• He usually asks 9-10 questions. Answer them and prompt him to edit the implementation-plan.md accordingly, so it's even better.

Your first implementation prompt

• Open Codex or Claude Code in VSCode's extensions or launch Claude Code or Codex CLI in the terminal of your project.

• Prompt: Read all the documents in /memory-bank, and proceed with Step 1 of the implementation plan. I will run the tests. Do not start Step 2 until I validate the tests. Once I validate them, open progress.md and document what you did for future developers. Then add any architectural insights to architecture.md to explain what each file does.

• Always start with "Ask" mode or "Plan Mode" (shift+tab in Claude Code) and once you are satisfied, allow the AI to go through the step.

• Extreme vibe: Install Superwhisper to speak casually with Claude or GPT-5 instead of typing.

Workflow

• After completing Step 1:
• Commit your changes to Git (if unfamiliar, ask your AI for help).
• Start a new chat (/new or /clear).
• Prompt: Now go through all files in the memory-bank, read progress.md to understand prior work, and proceed with Step 2. Do not start Step 3 until I validate the test.
• Repeat this process until the entire implementation-plan.md is complete.

---

Adding Details

Congratulations, you’ve built the base game! It might be rough and lack features, but now you can experiment and refine it.

• Want fog, post-processing, effects, or sounds? A better plane/car/castle? A gorgeous sky?
• For each major feature, create a new feature-implementation.md file with short steps and tests.
• Implement and test incrementally.

---

Fixing Bugs and Stuckness

• If a prompt fails or breaks the game:
• Use /rewind in Claude Code and refine your prompt until it works. If using GPT-5, you can commit often to git and reset when needed.
• For errors:
  • If JavaScript: Open the console (F12), copy the error, and paste it into VSCode to provide a screenshot for visual glitches.
  • Lazy Option: Install BrowserTools to skip manual copying/screenshotting.
• If stuck:
  • Revert to your last Git commit (git reset) and retry with new prompts.
• If really stuck:
  • Use RepoPrompt or uithub to get your whole codebase in one file and ask GPT-5 or Claude for assistance.

---

Claude Code & Codex Tips

• Codex CLI or Claude Code in the terminal: Run either tool inside VSCode's terminal to view diffs and feed additional context without leaving your workspace.
• Claude Code /rewind: Use this command to roll the project back to an earlier state if an iteration misses the mark.
• Custom Claude Code commands: Create helpers like /explain $arguments that trigger a prompt such as "Do a deep-dive on the code and understand how $arguments works. Once you understand it, let me know, and I will provide the task I have for you." so the model pulls in rich context before editing.
• Clearing context: Clear context frequently with /clear or /compact if you still need previous conversations context.
• Save time (at your own risk): Use claude --dangerously-skip-permissions or codex --yolo to start Claude Code or Codex CLI in a mode where it will never ask you confirmations.

Other Tips

• Small Edits: Use GPT-5 (medium)
• Great Marketing Copywriting: Use Opus 4.1
• Generate Great Sprites (2D images): Use ChatGPT and Nano Banana
• Generate Music: Use Suno
• Generate Sound Effects: Use ElevenLabs
• Generate Video: Use Sora 2
• Better prompt outputs:
  • Add “think as long as needed to get this right, I am not in a hurry. What matters is that you follow precisely what I ask you and execute it perfectly. Ask me questions if I am not precise enough."
  • For Claude Code, use specific phrases to trigger deeper reasoning: think < think hard < think harder < ultrathink.

---

Frequently Asked Questions

Q: I am making an app, not a game, is this the same workflow?
A: It's mostly the same workflow, yes! Instead of a GDD (Game Design Document), you can do a PRD (Product Requirements Document). You can also use great tools like v0, Lovable, or Bolt.new to prototype first and then move your code to GitHub, and then clone it to continue on VSCode or in the terminal with this guide.

Q: Your plane in your dogfight game is amazing, but I can’t replicate it in one prompt!
A: It’s not one prompt—it’s ~30 prompts, guided by a specific plane-implementation.md file. Use sharp, specific prompts like “cut out space in the wings for ailerons,” not vague ones like “make a plane.”

Q: Why is Claude Code or Codex CLI better than Cursor right now?
A: It really is up to your liking. We highlight that Claude Code is better at using Claude Sonnet 4.5, and Codex CLI is better at using GPT-5 than Cursor is at using either of them. Having them live in the terminal unlocks many more development workflows: working from any IDE, hopping onto a remote server through SSH, and so on. There are powerful customization options such as custom commands, sub-agents, and hooks that will speed up both the quality and the pace of development over time. Finally, if you’re on the lower-tier Claude or ChatGPT plan, that’s enough to get started.

Q: I don't know how to set up a server for my multiplayer game
A: Ask your AI.

---