I invented a new technique for Cursor. Using this, I reduced AI...
Prajwal Tomar@PrajwalTomar_
60 views
Mar 23, 2025
~2 min read
Advertisement
1
I invented a new technique for Cursor.
Using this, I reduced AI hallucinations and assumptions by 85%.
I am calling it “knowledge base” technique.
Here’s how it works: ↓
Using this, I reduced AI hallucinations and assumptions by 85%.
I am calling it “knowledge base” technique.
Here’s how it works: ↓
2
The golden rule for using AI in coding workflow.
You should never rely on AI to plan your codebase. That’s your job. Your job is to think and plan. The AI’s job is to implement the code.
You can force AI to just implement code by providing a detailed Knowledge base around your project.
Knowledge base includes your project docs.
Go to @CodeGuideDev and generate a set of project docs: 👇
You should never rely on AI to plan your codebase. That’s your job. Your job is to think and plan. The AI’s job is to implement the code.
You can force AI to just implement code by providing a detailed Knowledge base around your project.
Knowledge base includes your project docs.
Go to @CodeGuideDev and generate a set of project docs: 👇
3
1. PRD (Project Requirements Doc)
Start with this. It sets the tone.
Include:
• App overview
• User flows
• Tech stack & APIs
• Core features
• In-scope vs out-of-scope items
This gives the AI a solid high-level understanding of what you’re building.
Start with this. It sets the tone.
Include:
• App overview
• User flows
• Tech stack & APIs
• Core features
• In-scope vs out-of-scope items
This gives the AI a solid high-level understanding of what you’re building.
0:17
4
2. App Flow Doc
This is the map of your app. The clearer it is, the better AI performs.
Break down:
• Describe every page in your app
• How users move from one to another
• Simple language, no bullets
• Be painfully specific
Avoid vagueness. AI get confused easily.
This is the map of your app. The clearer it is, the better AI performs.
Break down:
• Describe every page in your app
• How users move from one to another
• Simple language, no bullets
• Be painfully specific
Avoid vagueness. AI get confused easily.
0:08
5
3. Tech Stack Doc
Tell the AI exactly what to build with.
List:
• All packages & dependencies
• Links to API docs (yes, it can read them)
• Preferred libraries or tools (e.g., Supabase, Stripe, NextAuth)
This keeps it from hallucinating random tech choices.
Tell the AI exactly what to build with.
List:
• All packages & dependencies
• Links to API docs (yes, it can read them)
• Preferred libraries or tools (e.g., Supabase, Stripe, NextAuth)
This keeps it from hallucinating random tech choices.
0:12
6
4. Frontend Guidelines
Give it your design system.
Include:
• Fonts
• Color palette
• Spacing & layout rules
• Preferred UI library or framework
• Icon set
If you want consistent design, you have to teach the AI your visual language.
Give it your design system.
Include:
• Fonts
• Color palette
• Spacing & layout rules
• Preferred UI library or framework
• Icon set
If you want consistent design, you have to teach the AI your visual language.
0:13
7
5. Backend Structure Doc
Especially important if you’re using Supabase or Firebase.
Tell AI:
• DB schema
• Auth logic
• Storage rules
• Any edge cases
AI can write SQL, but only if it knows what you want it to store.
These 5 docs alone will make your build 10x more stable.
But if you want to go pro, you need one more:
Especially important if you’re using Supabase or Firebase.
Tell AI:
• DB schema
• Auth logic
• Storage rules
• Any edge cases
AI can write SQL, but only if it knows what you want it to store.
These 5 docs alone will make your build 10x more stable.
But if you want to go pro, you need one more:
0:19
8
6. Implementation Plan
This is your anti-hallucination weapon.
List 50+ specific steps you’d take if you were coding this manually.
Give this to Cursor or Windsurf, and now they’re just executing, not guessing.
You can create all these docs using ChatGPT.
I use @CodeGuidedev to generate them automatically. It is built for AI coding models.
It uses GPT o1/03-mini, GPT-4o and Claude 3.5 models to generate structured docs.
Here’s a sample of my implementation plan:
This is your anti-hallucination weapon.
List 50+ specific steps you’d take if you were coding this manually.
Give this to Cursor or Windsurf, and now they’re just executing, not guessing.
You can create all these docs using ChatGPT.
I use @CodeGuidedev to generate them automatically. It is built for AI coding models.
It uses GPT o1/03-mini, GPT-4o and Claude 3.5 models to generate structured docs.
Here’s a sample of my implementation plan:
0:13
9
Takeaway:
• AI should implement, not think
• Your docs build the context boundary
• That’s how you go from chaos to clean builds
• Use tools like @CodeGuidedev or @ChatGPTapp to prep this FAST
• Then hand it off to Cursor or Windsurf for coding
AI can break your codebase but,
If you provide detailed context, it’ll blow your mind.
What’s your strategy to code production ready apps?
• AI should implement, not think
• Your docs build the context boundary
• That’s how you go from chaos to clean builds
• Use tools like @CodeGuidedev or @ChatGPTapp to prep this FAST
• Then hand it off to Cursor or Windsurf for coding
AI can break your codebase but,
If you provide detailed context, it’ll blow your mind.
What’s your strategy to code production ready apps?
10
Here’s another Cursor thread you might find helpful:
Prajwal Tomar@PrajwalTomar_
Still converting Figma designs manually?
Cursor + MCP now lets you convert real Figma designs into clean, responsive React components in seconds.
Here’s the exact workflow.
Cursor + MCP now lets you convert real Figma designs into clean, responsive React components in seconds.
Here’s the exact workflow.
