Hi,
I have been getting into using o1 for my recent projects, and I want to learn more about how prompting goes in o1 because from what I have seen from my tests are that it does not work well to the normal prompt’s that are used for gpt-4 and such, so just reaching out to see if anyone has any articles or anything they could share to help me learn more about o1 prompting 
Thank you
2 Likes
This is good question. I too wished to know the answer as I can “usually” get better answers out of 4o than o1, depending on the topic. So, I decided to ask o1 for what it advises, and it basically told me to keep using 4o until I have enough details and context built up in a thread before prompting o1 for a more reasoned answer. That’s basically what I’m already doing. Here is its response:
Thought about prompting adjustments for models for 8 seconds
When comparing prompts for ChatGPT o1 and GPT-4o, you generally don’t need to make significant changes to your overall prompting style. Both models respond well to clear, specific instructions and benefit from well-defined context, goals, and examples. That said, here are a few slight considerations when prompting ChatGPT o1 versus GPT-4o:
- Complex Reasoning and Depth
- GPT-4o typically handles more complex and nuanced queries with fewer details needed in the prompt. It can tolerate more ambiguity and still produce a coherent, in-depth response.
- ChatGPT o1 may sometimes benefit from a bit more explicit structure—enumerating steps, clarifying the exact type of answer you need, or including slightly more background information.
- Context Length and Compression
- GPT-4o often handles lengthy context more robustly and can maintain coherence over more extended conversations or multiple subtopics within the same prompt.
- ChatGPT o1 is still effective but can sometimes perform better if you break down especially long, multi-part questions into smaller, more targeted prompts.
- Iterative Refinement
- While both models benefit from iterative refinement (i.e., follow-up questions or instructions to steer them in the right direction), GPT-4o often requires fewer iterations to home in on highly accurate or deeply reasoned responses.
- With ChatGPT o1, providing follow-up prompts or clarifications can be equally effective, but you may need one or two extra steps for optimal detail and accuracy.
Overall, the same best practices—such as being clear, concise, and explicit with your request—apply to both ChatGPT o1 and GPT-4o. However, if you notice that ChatGPT o1 needs more clarity or detail in a particular domain (especially in highly specialized or intricate queries), refining your prompt to explicitly outline the needed steps or information will help ensure you get the best possible response.
2 Likes
Hi @programmerrdai !
My approach has been to make the instruction as short as possible and without using any Markdown formatting and such. Usually it’s 2-3 sentences max for o1. If I compare to my gpt-4o instructions (system prompts in this case), I would usually have it Markdown formatted, with a role definition followed by bulleted list of instructions with examples, and some “guidelines”. I found that if I leave this out from o1, it works much better.
1 Like