Post

More from @ihtesham2005

Ihtesham Ali

@ihtesham2005

Feb 22

🚨BREAKING: Someone leaked the system prompts of every major AI coding assistant.

You can see exactly how they work:

- Cursor, Claude Code, Windsurf, Devin, Replit, v0
- 30,000+ lines of internal instructions & prompts
- Agent architectures, tool configurations, workflows
- Lovable, Perplexity, NotionAI implementations
- 30+ AI tools completely reverse-engineered

116k stars. 100% Opensource.

Repo: github.com/x1xhlol/system…

What's inside:

→ Exact system prompts from $20/month AI coding tools
→ Internal tool configurations
→ Model selection strategies
→ Prompt engineering patterns they use
→ Complete architecture of 30+ AI products

You're literally seeing the source code of how these tools work.

Read 4 tweets

Ihtesham Ali

@ihtesham2005

Feb 19

If you're serious about AI, learn these 10 concepts ASAP:

→ Context windows
→ Agent memory
→ Tool use
→ RLHF
→ Distillation
→ Evaluation
→ Prompt chaining
→ Self-critique
→ Retrieval
→ Multi-agent systems

Most people ignore them. Here's why each one matters 👇

1. Context windows

This is how much text a model can "see" at once.

GPT-4o: 128K tokens
Gemini 2.5 Pro: 1M tokens
Claude: 200K tokens

Bigger isn't always better what you PUT in the window matters more than the size.

Most people fill it with garbage and wonder why outputs suck.

2. Agent memory

LLMs forget everything the second a conversation ends.

Memory fixes that.

3 types:

- In-context (short-term, inside the prompt)
- External (databases, vector stores)
- Episodic (summarized past sessions)

Without memory, you're building agents that are goldfish.

Read 12 tweets

Ihtesham Ali

@ihtesham2005

Feb 9

I spent 3 weeks analyzing the most powerful ChatGPT research prompts that actual academics are hiding.

The difference between spending 6 hours on literature review vs 8 minutes.

12 prompts I use daily for my PhD work.

Steal them 👇

1. THE LITERATURE SYNTHESIZER

"I'm researching [topic]. Synthesize the key arguments from these 5 papers: [paste abstracts]. Identify theoretical frameworks, methodology gaps, and conflicting findings. Create a comparison table."

Turns 3 hours of note-taking into 4 minutes.

2. THE GAP FINDER

"Based on this literature review [paste your summary], identify 3-5 unexplored research questions that would be publishable in top-tier journals. For each gap, explain why it matters and suggest a methodology."

This is how you find dissertation topics in 10 minutes.

Read 15 tweets

Ihtesham Ali

@ihtesham2005

Feb 6

10 powerful claude opus 4.6 prompts that feel illegal to know 👇

1. Code Architect

Stop getting garbage code. This prompt makes Claude write production-ready functions.

Mega prompt:

You are a senior software architect. Write [FUNCTIONALITY] with these requirements:

Technical specs:
- Language: [LANGUAGE/FRAMEWORK]
- Performance: [REQUIREMENTS]
- Error handling: [SPECIFICATIONS]
- Testing: Include unit tests

Standards:
- Follow [STYLE GUIDE]
- Document all functions
- Optimize for [METRIC]
- Handle edge cases: [LIST]

Provide:
1. Complete implementation
2. Usage examples
3. Performance considerations
4. Potential issues and solutions

Context: [YOUR CODEBASE DETAILS]

2. Content Strategist

Generate content calendars that actually convert, not generic bullshit.

Mega prompt:

You are a content strategist who's generated $10M+ in revenue through content.

Create a 30-day content calendar for [PLATFORM] targeting [AUDIENCE].

For each piece provide:
1. Topic and angle
2. Hook/headline (3 variations)
3. Key points to cover
4. SEO keywords
5. CTAs that convert
6. Best posting time/day

Goals: [METRICS - traffic, leads, sales]
Brand voice: [DESCRIPTION]
Avoid: [TOPICS/ANGLES]

Include rationale for each piece based on funnel stage.

Read 12 tweets

Ihtesham Ali

@ihtesham2005

Feb 5

Founders and AI prompt engineers leaked these advanced LLM workflows that separate novices from experts.

I've been applying insider insights from top AI builders for 8 months. The impact is unreal.

Here are 8 prompt engineering patterns they don't want you to know (but I'm sharing anyway):

1. Chain-of-Verification (CoVe)

Most people ask once and accept the answer.

Experts force the model to verify its own output.

Prompt structure:

- Generate initial response
- "Now list assumptions you made"
- "Verify each assumption"
- "Revise based on verification"

Reduces hallucinations by 40%.

2. Constitutional AI Prompting

Instead of telling the model what TO do, define what NOT to do.

"Your response must not:

- Include speculation without labeling it
- Cite sources you can't verify
- Use jargon without explanation"

This is how Anthropic trains Claude internally.

Read 11 tweets

Ihtesham Ali

@ihtesham2005

Feb 4

I scraped every single NotebookLM prompt that blew up on X, Reddit, and academic corners of the internet.

Turns out most people are using NotebookLM like a fancy note-taker.

That's insane.

It's a full-blown research assistant that can compress 10 hours of analysis into 20 seconds if you feed it the right instructions.

Here's what actually works:

Prompt 1: The Expert Synthesizer

"You are a [field] expert with 15 years of experience. Analyze these sources and identify the 3 core insights that practitioners in this field would immediately recognize as groundbreaking. For each insight, explain why it matters and what conventional wisdom it challenges."

This forces depth over breadth. The output is immediately usable.

Prompt 2: The Contradiction Hunter

"Compare these sources and identify every point where they contradict each other. For each contradiction, explain which source has stronger evidence and why. If both are credible, explain what factors might explain the disagreement."

Perfect for literature reviews and due diligence. Saves hours of manual cross-referencing.

Read 14 tweets

Share this page!

Enter URL or ID to Unroll

Ihtesham Ali

Try unrolling a thread yourself!

More from @ihtesham2005

Ihtesham Ali

Ihtesham Ali

Ihtesham Ali

Ihtesham Ali

Ihtesham Ali

Ihtesham Ali

Did Thread Reader help you today?

Don't want to be a Premium member but still want to support us?

Send Email!