Wanna vibe code better products? Here are some tips and tricks I can share after 1800 commits and more than 1000 hours of vibe coding: 1. Start by explaining your idea to the LLM and get it to prepare you a PRD and architecture documents (as md files). 2. Load the md files into your LLM, let it read them all, and tell it to create an implementation plan document. 3. To get the most out of it, instead of using a general LLM, have subagents with specific skills take on different parts. This will give you a set of specialized team members, instead of one generic LLM, that's jack of all trades. 4. For number 3 to be effective, tell your LLM to code the agents as world-class in their domain, trained by the best in-class people, or even better, use a tool to spawn the agents (I’ll be sharing a tool I built for this soon, called Spawner). 5. During implementation, make sure you’re using GitHub commits as much as you can. Instead of trying to one-shot the whole app in one go, test features to make sure each one is working before you move on to the next. 6. For UI design not to look like slop, I highly recommend either importing references or using a UI-skilled agent with more context about: (1) how you want users to experience your product, (2) who your target audience is, (3) get one page looking amazing, (4) let the AI use that styling on the rest of your product, and (5) tell AI to avoid gradients and bad icons because it usually defaults to those. 7. AI is not great at building authorization systems by itself, especially for X. Use a skilled agent for this too: it’ll make everything a lot easier. You can spend hours pulling your hair out if you don’t. 8. If you don’t like certain features and want to change them, make sure afterward you tell AI to review your codebase and clean up any unused features and code. Even better: go back to commits before a certain feature was coded and build the version you want from there. If you let your code become spaghetti, it’ll cause so much headache you’ll want to abandon the project: it becomes unbearable to fix things, even with AI. 9. Tell AI to use a changelog document and make sure your initial architecture/PRD documents are updated to reflect your current features consistently. AI may get a wrong impression of what your featureset looks like if your initial docs and current features diverge, which can lead to spaghetti features and code. 10. Use vulnerability scanners to check whether you’ve left secrets exposed or have security gaps: AI tends to have these issues. I built a tool for this too; check vibeship dot co. Good luck!

Wanna vibe code better products? Here are some tips and tricks I can share after 1800 commits and more than 1000 hours of vibe coding: 1. Start by explaining your idea to the LLM and get it to prepare you a PRD and architecture documents (as md files). 2. Load the md files into your LLM, let it read them all, and tell it to create an implementation plan document. 3. To get the most out of it, instead of using a general LLM, have subagents with specific skills take on different parts. This will give you a set of specialized team members, instead of one generic LLM, that's jack of all trades. 4. For number 3 to be effective, tell your LLM to code the agents as world-class in their domain, trained by the best in-class people, or even better, use a tool to spawn the agents (I’ll be sharing a tool I built for this soon, called Spawner). 5. During implementation, make sure you’re using GitHub commits as much as you can. Instead of trying to one-shot the whole app in one go, test features to make sure each one is working before you move on to the next. 6. For UI design not to look like slop, I highly recommend either importing references or using a UI-skilled agent with more context about: (1) how you want users to experience your product, (2) who your target audience is, (3) get one page looking amazing, (4) let the AI use that styling on the rest of your product, and (5) tell AI to avoid gradients and bad icons because it usually defaults to those. 7. AI is not great at building authorization systems by itself, especially for X. Use a skilled agent for this too: it’ll make everything a lot easier. You can spend hours pulling your hair out if you don’t. 8. If you don’t like certain features and want to change them, make sure afterward you tell AI to review your codebase and clean up any unused features and code. Even better: go back to commits before a certain feature was coded and build the version you want from there. If you let your code become spaghetti, it’ll cause so much headache you’ll want to abandon the project: it becomes unbearable to fix things, even with AI. 9. Tell AI to use a changelog document and make sure your initial architecture/PRD documents are updated to reflect your current features consistently. AI may get a wrong impression of what your featureset looks like if your initial docs and current features diverge, which can lead to spaghetti features and code. 10. Use vulnerability scanners to check whether you’ve left secrets exposed or have security gaps: AI tends to have these issues. I built a tool for this too; check vibeship dot co. Good luck!