Author: Nicolas Zullo, https://x.com/NicolasZu
Creation Date: March 12, 2025
Last Update Date: May 25, 2026
To begin vibe coding, you only need one of these two tools:
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). Right now, the Codex App is gaining a lot of love.
(Note: While earlier versions of this guide utilized Grok 3, we then transitioned to Gemini 2.5 Pro. And now we're using Claude Opus 4.7 (or Codex 5.5))
(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.
npm i -g @openai/codexnpm i -g @anthropic-ai/claude-code
(or download the Codex app)game-design-document.md.AGENTS.mdtech-stack.md.
/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.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.
game-design-document.md)tech-stack.md).md) which is a set of step-by-step instructions for your AI developers.
memory-bank.memory-bank:
game-design-document.mdtech-stack.mdimplementation-plan.mdprogress.md (Create this empty file for tracking completed steps)architecture.md (Create this empty file for documenting file purposes)Now the fun begins!
/memory-bank, is implementation-plan.md clear? What are your questions to make it 100% clear for you?implementation-plan.md accordingly, so it's even better./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.shift+tab in Claude Code) and once you are satisfied, allow the AI to go through the step./new or /clear). Why? LLM produce their best results when there is a lot of room in their context window.implementation-plan.md is complete.Congratulations, you’ve built the base game! It might be rough and lack features, but now you can experiment and refine it.
feature-implementation.md file with short steps and tests./rewind in Claude Code and refine your prompt until it works. If using Codex 5.5, you can commit often to git and reset when needed.F12), copy the error, and paste it into VSCode to provide a screenshot for visual glitches.git reset) and retry with new prompts./rewind: Use this command to roll the project back to an earlier state if an iteration misses the mark./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./clear or /compact if you still need previous conversations context. Keep /context over 50 or 60% for best performances.claude --dangerously-skip-permissions or codex --yolo to start Claude Code or Codex CLI in a mode where it will never ask you confirmations.think < think hard < think harder < ultrathink.Q: What projects have been 100% vibe coded by yourself with this methodology?
A: Among the most recent:
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 Opus 4.7, and Codex CLI is better at using Codex 5.5 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.
115 followers · starred Mar 2025
174 followers · starred May 2025
85 followers · starred Dec 2025
88 followers · starred Apr 2025
Author: Nicolas Zullo, https://x.com/NicolasZu
Creation Date: March 12, 2025
Last Update Date: May 25, 2026
To begin vibe coding, you only need one of these two tools:
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). Right now, the Codex App is gaining a lot of love.
(Note: While earlier versions of this guide utilized Grok 3, we then transitioned to Gemini 2.5 Pro. And now we're using Claude Opus 4.7 (or Codex 5.5))
(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.
npm i -g @openai/codexnpm i -g @anthropic-ai/claude-code
(or download the Codex app)game-design-document.md.AGENTS.mdtech-stack.md.
/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.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.
game-design-document.md)tech-stack.md).md) which is a set of step-by-step instructions for your AI developers.
memory-bank.memory-bank:
game-design-document.mdtech-stack.mdimplementation-plan.mdprogress.md (Create this empty file for tracking completed steps)architecture.md (Create this empty file for documenting file purposes)Now the fun begins!
/memory-bank, is implementation-plan.md clear? What are your questions to make it 100% clear for you?implementation-plan.md accordingly, so it's even better./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.shift+tab in Claude Code) and once you are satisfied, allow the AI to go through the step./new or /clear). Why? LLM produce their best results when there is a lot of room in their context window.implementation-plan.md is complete.Congratulations, you’ve built the base game! It might be rough and lack features, but now you can experiment and refine it.
feature-implementation.md file with short steps and tests./rewind in Claude Code and refine your prompt until it works. If using Codex 5.5, you can commit often to git and reset when needed.F12), copy the error, and paste it into VSCode to provide a screenshot for visual glitches.git reset) and retry with new prompts./rewind: Use this command to roll the project back to an earlier state if an iteration misses the mark./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./clear or /compact if you still need previous conversations context. Keep /context over 50 or 60% for best performances.claude --dangerously-skip-permissions or codex --yolo to start Claude Code or Codex CLI in a mode where it will never ask you confirmations.think < think hard < think harder < ultrathink.Q: What projects have been 100% vibe coded by yourself with this methodology?
A: Among the most recent:
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 Opus 4.7, and Codex CLI is better at using Codex 5.5 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.
115 followers · starred Mar 2025
174 followers · starred May 2025
85 followers · starred Dec 2025
88 followers · starred Apr 2025