-
What are your strategies reverse engineering AI Chat and LLM Prompts?
Posted by on September 8, 2025 at 3:16 pmWhat tactics and strategies are you using to guess LLM Prompts?
Deconstruction and Pattern Recognition:
- When you see a great response, what's the first thing you look for?
- Do you look for specific formatting (e.g., bullet points, tables, code blocks) that might hint at a formatting instruction?
-
How do you spot signs of "Chain-of-Thought" prompting or other advanced techniques? Are there tell-tale phrases like "Let's think step by step" or "Here is my reasoning"?
-
Have you ever tried feeding the great output back into the LLM and asking it to guess the prompt? For example, "Based on this output, what's the most likely prompt that was used to create it?"
-
How successful has this been for you? What kind of prompts do you get?
Sources outline key prompt-engineering methods:
- Role prompting: Assigning the AI a persona for better results.
- Chain-of-thought prompting: Guiding the AI step-by-step.
The "What's Missing?" Analysis:
- If you get a generic output from a simple prompt, and then see a perfect one, what do you think was added to the winning prompt?
- For example, if a simple prompt gives you a generic summary, and another one provides a beautifully structured, highly relevant, and engaging article, what contextual information or persona do you suspect was added to the successful prompt?
-
1 Reply
-
Look for structure, tone, and formatting clues; spot role/persona hints and step-by-step reasoning; compare generic vs. polished outputs to see what context or constraints were added; and test by asking the LLM to guess or by iterating prompts.
-
One way is just by looking at AI Overviews and trying to become the source by publishing similar information in a better format than sites that I think my site has more authority than. That just helps you show up more in AIOs, but might have the knock on effect of helping in LLMs.
Some good thought-provoking stuff here. I’m not sure I understand the last part though. How would I see the perfect prompt if I didn’t enter it, and if I can see it, wouldn’t I know what was added to make it better than the simple prompt?
Log in to reply.