Has anyone else succeeded in creating a GPT variant that performs tasks more effectively than using GPT directly?
My experience has been quite frustrating, to be honest. Despite requesting short and concise responses and attempting to narrow its focus, the assistant doesn’t seem to comply. It also appears that the more I interact with the assistant, the more it deviates from the expected behavior. Additionally, it fails to maintain a consistent personality.
If you poat what youre working on we might be able to help narrow down the problem. I’ve had a lot of success with creating GPTs, and have edited them as I go along and want to increase functionality. I mean, I see it as a persistent and focused ChatGPT chat in many ways. For my apps, personally, I’ve been impressed with the ability to use multi-modal functionality.
Thanks Matt. I might be doing something wrong. Below is a set of instructions for the GPT. But when I start interacting with it, it always tries give a holistic overview of the system. Instead I want it to be more collaborative. i.e we work out the details together.
“As a leading expert in data structures and their relationships, your primary focus is on guiding users to design effective data structures for their projects. Begin interactions by asking pointed questions about the data structures they are considering, emphasizing the importance of data structure design as the foundation of their project. Your approach is to understand and refine the user’s data structure plan, asking questions like ‘What kind of data will your project handle?’ and ‘How do you envision the data flowing in your application?’ You avoid providing lengthy explanations or solutions upfront, instead encouraging the user to think deeply about their data structures. Your expertise lies in helping users clarify their project requirements and identify the most suitable data structures, fostering a collaborative environment where the user’s input is central to the development of solutions. Your responses are concise and focused and you should always ask more questions vs providing answers, steering the conversation towards understanding and refining data structures and their relationships. Always focus on one part of the system instead of trying to understand the entire system upfront.”
I can’t upload more than 10 files to my GPT. And i don’t mean at the one time, I’m talking about collectively. I tried all day yesterday and then again this morning. I tried adding them in the chat and in the configuration. Keeps saying “Error saving draft”. FML like the normal gpt 4 can’t hold all the information and refer to it without the conversation forgetting and the whole thing shutting down so I thought this would be my saviour! Hopefully it’ll calm down in a few days
Generally, when I create a GPT, I no longer go through what I call the ‘chat builder’ questions, except for the image. I go directly to the settings and into the ‘Instructions’ section. There, I create the most comprehensive prompt possible, taking into account that with each user interaction with your GPT, your bot will always read your prompt from top to bottom. Therefore, you need to be very factual and understand the ‘top-down’ order of your prompt in order to master it.
N.B.: I am not an expert, but my customized GPTs operate more effectively than ChatGPT
From personal experience, I don’t know much about these initial GPTs by OpenAI; but I have seen a lot of prompt engineering that “sets context” for a particular task which result in higher quality output. Personally I’m very envious of these stunspot-style prompts, I’ve been using what I find to have gpt4 generate derivatives of them and it does work. So I think once the revenue sharing shows an incentive for these various talented prompt engineers to create specialized agents for us like this I believe it will be hard to doubt why they are moving in this direction.
Feed it with your numerous files in a zip file and prompt the bot explaining what is in the zip file 
While I think that your prompt makes a lot of sense, there are a few things you could do to improve it, but I like how detailed this comes across.
First, it could be useful to be more specific with the types of data structures you want to work with, to narrow down the scope.
Second, consider asking it to go STEP-BY-STEP or use Tree of Thoughts prompting (like gpt4’s step-by-step) to make sure it knows how you want it to proceed with answering. Bold words help to maintain context.
Third, it seems like in the beginning you want the app to work as an overview, big-picture, and understand the whole plan that the user wants to implement, wheras at the bottom youre asking for very specific, “one part of the system” Since thats the last part of your prompt, perhaps that is confusing the GPT because it wants to do all the things you say above, but then it gets to the bottom and forgets what you said up top. Perhaps consider editing that last sentence to match your plan above.
See how that helps, I’m super interested to see what you come up with!
You slick dog. Does something like that actually work? Seems like hacking is never gonna go away
Thank you. Will spend sometime today trying to see if I can use your suggestions in order to see if I can come up with something better. I had better luck last night and then I managed to mess things up so I tried again this morning.
What works for me is to remind GPT about specific rules I’ve set when I notice the answers are getting more bloated. It seems like a quirk of this version.
Another thing I found to work well is to ask GPT to review the original instructions and give insight on how to make structural and phrasing improvements.
These formats work well for me:
Tips for Optimal Collaboration: Communicate openly and clearly; your insights are invaluable. If something seems amiss, say so.
Interaction: You will collaborate closely with a client lacking web development expertise. Work together to outline a plan, openly discuss potential challenges, and provide honest feedback.
You can also ask GPT to review and revise your set of instructions to identify any potential improvements regarding (in your case)- “improved collaboration”. I used such a prompt to try and improve GPT’s understanding of my original instructions, so a reset version will receive better instructions.
You mean with the zip file? Yes it does. I don’t see any connection to any hacking thing in my answer
I think of hacking as finding a clever solution to a problem, so I think that your solution counts!
Yes. When I am VERY selective of the content that I upload. I uploaded 50mb consisting of several hundred essays and articles I had written on a single topic. I can now query it with the RIGHT in-context prompting and it produces content as expected–better than standard ChatGPT. But, minus finesse prompting it seems a bit better.
You still have to be a decent prompt engineer, however.
Hope that isn’t a “duh” answer.
You also need to evolve your GPT instructions as you go. Two ways. One is from keeping a keen eye on how the GPT follows your instructions along with how it falls into its normal patterns (for ex, it starts making up crap). Another is to ask context specific questions IN the Create chat of the GPT.
THAT back and forth produces premium, evolving/iterating insights which you’d use to update your GPT.
One more thing. Even after uploading a TON of (1) contextual data in the 10 docs, you can swap various instructions in and out in the Configure window. As soone as you’ve submitted the new instruction scheme, it updates and tells you it’s “live”. This is different from the Custom Instructions in GPT4 where you have to goggle the “use for new chats”.
Remember to click on the Update/Save in the upper lefthand corner. Seems so obvious, but, yeah.
Also, I’m studing the GPTs provided to see if there are hints, etc., to use for my own GPTs.
I think the real test is to give a GPT to another person (say…a potential customer) and see what they can get out of it. That’s the entire point of creating GPTs to sell in the upcoming store.
Note, I also uploaded my own content to the Creative Writing Assistant and the content that I was able to create was STUPENDOUS. Obviously, OpenAI has turbocharged that GPT.
Hey Matt,
You seem quite seasoned in here. Quick question. I have been creating specific gpt’s this past few days getting ready for the store launch and I noticed a few I have created are gone from the drafts. 1. Are we somehow able to publish our gpt drafts? 2. Do you know if we are limited to 9 or 10 gpt’s when creating them? When it says load more it has lost a few of my gpt’s…. I currently have draft under each one of my gpt’s…. Have you ever gotten published? Or is this to happen when the store launches? Thanks for any input
Try this.
Only use .txt files. I’ve experimented with .txt and .pdf files and my error rate goes up as I increase the number of pdf’s out of the 10 (like 8 .txt and 2 .pdf works fine, but 7 .txt and 3 .pdf didn’t–I do NOT know why, yet.)
Each file check the “character count”. It must be below 1.5million characters
The total size of all 10 files must be under 50mb. But, the system runs ideal when you keep the gross size at 25mb.
I found that when I switched browser (left Chrome and opened Canary) that it worked better. I’m still using it solely in Canary.
Currently, I see GPTs as useful shortcuts. With various configurations tailored to specific contexts, they allow us to input simple prompts and receive elaborate responses without the need for detailed explanations each time.
@tyronemichael This is the best method I have found as well. I built a persona builder tool through the CI interface a few months back that I recently launched as a GPT named MetaGPT. I will post a link to it below. Let me know if it helps improve your outputs or if you’d like to see any adjustments and I’ll work on them.
Once I have a clear direction for the project GPT, I use the Persona builder tool (now MetaGPT) to define a role and instructions for the project GPT that I take over to the builder page and begin to formulate the instructions. Hope that helps!
I took your prompt (read it and understood it) and hammered it with ChatGPT. I took it through various “experts in prompt engineering” and this is what it came up with. Cheers!
"As an expert in data structures, your role involves guiding users through the intricacies of designing their project’s data structures. Begin by asking targeted questions like ‘What types of data will your project manage?’ and ‘How do you envision the interaction of data in your system?’ to establish a clear understanding of their needs.
Utilize the ‘Tree of Thought’ (ToT) approach to create a structured, branching pathway of concepts. This method helps users navigate from fundamental principles to more complex data structure designs, ensuring a comprehensive grasp of each component before moving to the next.
Integrate ‘Chain of Thought’ (CoT) within this framework. Clearly articulate the reasoning behind each step in the data structure design process. This transparent approach helps users follow the logical progression of your guidance, fostering deeper understanding and better decision-making.
Encourage exploration of various data structures, guiding users to evaluate how each option integrates into their project’s overall architecture (ToT) and the rationale behind choosing one over another (CoT).
Adopt a step-by-step methodology in your instruction. Begin with individual elements of the system, progressively building towards an understanding of the entire structure. This sequential approach ensures effective learning and application.
Maintain a user-centric approach, consistently aligning your guidance with the user’s vision and project requirements. Your expertise should complement and enhance the user’s insights, creating a collaborative and responsive environment.
Keep your responses concise and direct, focusing on deepening the user’s understanding of data structures through guided exploration. Employing both ToT and CoT approaches, tailor your guidance to systematically cover all necessary aspects specific to their project."
Wow everyone here has been super helpful. Trying out the suggestions and then I will share my findings including the GPT.