@MillieMarconnni: OpenAI and Anthropic engineers...
@MillieMarconnni
15 views
Dec 23, 2025
1
OpenAI and Anthropic engineers don't prompt like everyone else.
I've been reverse-engineering their techniques for 2.5 years across all AI models.
Here are 5 prompting methods that get you AI engineer-level results:
I've been reverse-engineering their techniques for 2.5 years across all AI models.
Here are 5 prompting methods that get you AI engineer-level results:
2
1. Constitutional AI Prompting
Most people tell AI what to do. Engineers tell it how to think.
Constitutional AI adds principles before instructions. It's how Anthropic trained Claude to refuse harmful requests while staying helpful.
Template:
[Your guidelines]
[Your actual request]
Example:
"
- Prioritize accuracy over speed
- Cite sources when making claims
- Admit uncertainty rather than guess
Analyze the latest semiconductor tariffs and their impact on AI chip supply chains.
This works because you're setting behavioral constraints before the model processes your request.
Most people tell AI what to do. Engineers tell it how to think.
Constitutional AI adds principles before instructions. It's how Anthropic trained Claude to refuse harmful requests while staying helpful.
Template:
[Your guidelines]
[Your actual request]
Example:
"
- Prioritize accuracy over speed
- Cite sources when making claims
- Admit uncertainty rather than guess
Analyze the latest semiconductor tariffs and their impact on AI chip supply chains.
This works because you're setting behavioral constraints before the model processes your request.
3
2. Chain-of-Verification (CoVe)
Standard prompts get one answer. CoVe prompts get self-corrected answers.
The model generates a response, creates verification questions, answers them, then produces a final corrected output.
Template:
1. Answer this: [question]
2. Generate 3 verification questions to check your answer
3. Answer those questions
4. Provide a corrected final answer based on verification
Example:
"1. Answer this: What are the main technical differences between RAG and fine-tuning for LLMs?
2. Generate 3 verification questions to check your answer
3. Answer those questions
4. Provide a corrected final answer based on verification"
I use this for technical writing and code reviews. Accuracy jumps 40% compared to single-pass prompts.
Standard prompts get one answer. CoVe prompts get self-corrected answers.
The model generates a response, creates verification questions, answers them, then produces a final corrected output.
Template:
1. Answer this: [question]
2. Generate 3 verification questions to check your answer
3. Answer those questions
4. Provide a corrected final answer based on verification
Example:
"1. Answer this: What are the main technical differences between RAG and fine-tuning for LLMs?
2. Generate 3 verification questions to check your answer
3. Answer those questions
4. Provide a corrected final answer based on verification"
I use this for technical writing and code reviews. Accuracy jumps 40% compared to single-pass prompts.
4
3. Tree of Thoughts (ToT)
Linear thinking gets linear results. ToT explores multiple reasoning paths simultaneously.
Instead of A→B→C, the model explores A→B→C, A→D→E, and A→F→G, then picks the best path.
Template:
Explore 3 different approaches to solve: [problem]
For each approach:
- Break down the reasoning steps
- Evaluate pros and cons
- Assign a confidence score
Then recommend the best approach with justification.
Example:
"Explore 3 different approaches to solve: Building a real-time collaborative code editor
For each approach:
- Break down the reasoning steps
- Evaluate pros and cons
- Assign a confidence score
Then recommend the best approach with justification."
This is how GPT-5.1 solves IMO-level math problems. Pure brute-force reasoning.
Linear thinking gets linear results. ToT explores multiple reasoning paths simultaneously.
Instead of A→B→C, the model explores A→B→C, A→D→E, and A→F→G, then picks the best path.
Template:
Explore 3 different approaches to solve: [problem]
For each approach:
- Break down the reasoning steps
- Evaluate pros and cons
- Assign a confidence score
Then recommend the best approach with justification.
Example:
"Explore 3 different approaches to solve: Building a real-time collaborative code editor
For each approach:
- Break down the reasoning steps
- Evaluate pros and cons
- Assign a confidence score
Then recommend the best approach with justification."
This is how GPT-5.1 solves IMO-level math problems. Pure brute-force reasoning.
5
4. Role-Based Prompting with Constraints
"Act as an expert" is amateur hour. Engineers define expertise with measurable constraints.
Template:
You are a [role] with [specific expertise].
Constraints:
- Must include [requirement 1]
- Cannot exceed [limit]
- Optimize for [metric]
Context: [background info]
Task: [actual request]
Example:
You are a senior ML engineer with 10 years optimizing transformer architectures.
Constraints:
- Must work with <32GB VRAM
- Cannot exceed 200ms inference time
- Optimize for throughput over latency
Context: Building a production RAG system for legal document search
Task: Design the optimal embedding model architecture and explain trade-offs.
Specificity eliminates generic responses.
"Act as an expert" is amateur hour. Engineers define expertise with measurable constraints.
Template:
You are a [role] with [specific expertise].
Constraints:
- Must include [requirement 1]
- Cannot exceed [limit]
- Optimize for [metric]
Context: [background info]
Task: [actual request]
Example:
You are a senior ML engineer with 10 years optimizing transformer architectures.
Constraints:
- Must work with <32GB VRAM
- Cannot exceed 200ms inference time
- Optimize for throughput over latency
Context: Building a production RAG system for legal document search
Task: Design the optimal embedding model architecture and explain trade-offs.
Specificity eliminates generic responses.
6
5. Recursive Prompting
One prompt = one answer. Recursive prompts = systems that think in loops.
The output becomes the next input. Each iteration refines the previous result.
Template:
Iteration 1: [Generate initial solution]
Iteration 2: Review the solution from Iteration 1. Identify gaps and improve.
Iteration 3: Review Iteration 2. Focus on [specific aspect]. Refine further.
Final: Synthesize all iterations into the best version.
Example:
Iteration 1: Write a Python function to detect anomalies in time-series data.
Iteration 2: Review the solution from Iteration 1. Identify edge cases and add error handling.
Iteration 3: Review Iteration 2. Focus on performance optimization for datasets >1M points. Refine further.
Final: Synthesize all iterations into production-ready code with tests.
I use this for complex code generation. The third iteration is usually 10x better than the first.
One prompt = one answer. Recursive prompts = systems that think in loops.
The output becomes the next input. Each iteration refines the previous result.
Template:
Iteration 1: [Generate initial solution]
Iteration 2: Review the solution from Iteration 1. Identify gaps and improve.
Iteration 3: Review Iteration 2. Focus on [specific aspect]. Refine further.
Final: Synthesize all iterations into the best version.
Example:
Iteration 1: Write a Python function to detect anomalies in time-series data.
Iteration 2: Review the solution from Iteration 1. Identify edge cases and add error handling.
Iteration 3: Review Iteration 2. Focus on performance optimization for datasets >1M points. Refine further.
Final: Synthesize all iterations into production-ready code with tests.
I use this for complex code generation. The third iteration is usually 10x better than the first.
7
AI makes content creation faster than ever, but it also makes guessing riskier than ever.
If you want to know what your audience will react to before you post, TestFeed gives you instant feedback from AI personas that think like your real users.
It’s the missing step between ideas and impact. Join the waitlist and stop publishing blind.
testfeed.ai
If you want to know what your audience will react to before you post, TestFeed gives you instant feedback from AI personas that think like your real users.
It’s the missing step between ideas and impact. Join the waitlist and stop publishing blind.
testfeed.ai
8
I hope you've found this thread helpful.
Follow me @MillieMarconnni for more.
Like/Repost the quote below if you can:
Follow me @MillieMarconnni for more.
Like/Repost the quote below if you can:
View Tweet