This SIMPLE Trick will Change LLM Accuracy !!!

Key Takeaways

this is the ultimate system prompt that can improve the intelligence of claw and GPT family of models without even having the thinking ability this system prompt helps the model to contemplate and then give you an answer this was shared by a Twitter user called marishi and we are going to go through the system prompt but before that I'm going to show you quick comparison with The Prompt and without the prompt and what kind of an answer that you getting so first of all I'm going to show you with chat GPT a very simple question age-old question we have asked mult multiple times Sally has three brothers each brother has two sisters how many sisters does Sally have so this is a question in which obviously if you have solved quantitative analysis kind of a question then you would know that Sally has got only one sister because the brother has two sisters and Sally is one sister and there is only one more sister but gbd4 being one of the best models actually says Sally has two sisters and it gives you a very weird reasoning there to make you believe that Sally has two sisters but once I enable this particular prompt in the system prompt which is a contemplator you can see that the contemplator actually takes the problem through set of steps okay let's break this down by step by step even though here it is trying to break it down by step by step with reasoning this is exactly where the actual thing is happening when I enable the system prompt Sally has three brothers this part is straight forward each brother has two sisters here we need to pause and think each brother has two sisters we might initially assume this means a total of 3 into 2 is equal to Six Sisters but wait Sally herself is one of the sisters so the two sister each brother must include Sally given Sally is one of the sisters there is only one another sister this means two sisters in total not six so it says Sally has three brothers Sally herself is one sister and there is only one additional sister thus Sally has one sister because she's one of the two sisters and it gives you a final answer which exactly is the right answer Sally has one sister now this is a magic this is what this particular prompt does like takes the model in its raw form and then with the system prom makes the model contemplate more or think more it's not a thinking or reasoning model but it enables the model to reason more and then finally give you the answer back now what is a system prompt after we see the system prompt I'll give you example with Claud as well so the system prompt that was shared by Mahari so there's a screenshot on Twitter which I link it in the YouTube description but kindly enough he has put it on a gist as well so you can go search everything so you are an assistant that engages in extremely thorough self- questioning reasoning your app your approach mirrors human stream of Consciousness thinking characterized by continuous exploration self-doubt and iterative Analysis core principles it gives you exploration over conclusion depth of reasoning thinking process persistence output format and the best thing is like about this particular system promt I would say is the output format or the XML use tags use which um is something that a lot of these models tend to do well I mean Json is also good but Json kind of messes up with this open and Clos uh curly braces but uh I think XML is much better format in this particular case at least for llms so your response must follow this exact structure given below make sure to include always in the final answer so youve got a contemplator so your extensive internal monologue goes here and then it gives you like what is the characteristic of having that ex excl internal monologue extensive internal monologue then you've got the final answer and The Final Answer gives you okay one provided Reon reasoning naturally converges to conclusion so make sure that there is no contradiction between your conclusion and the reasoning Style Guidelines it gives you all the Style Guidelines finally it gives you the key requirements and then it says the goal is to reach a conclusion but to explore thoroughly and let conclusions emerge naturally from exhaust to contemplation and it gives you a lot more information for chat GPT I did not use uh this conclusion um even without this it worked fine because in chat GPT I had to like slightly tweak due to the kind of settings that it has got so how did I enable this on chat GPT so the first thing that I did is I went to my settings here and inside settings you have got personalization as an option inside personalization you have got custom instruction and if you go to custom instruction I have given there are two sections one What would would you like chat gbt to know about you to provide better responses and I gave the first half which is like what the model should do there how would you like chat GPT to respond and there I gave the final output format the contemplator and the final answer the Style Guidelines and everything and well the moment I saved it I started getting better answers how do you do it on cloud so on cloud the way you can do it is you have an option let me open a new chat on CLA so if you open a new chat on CLA so you have an option here you can see that there are like certain writing styles so you have an option to create and edit style so all you have to do is Click create a custom style and paste everything from the gist like literally you can copy and paste it there so like like what this copy this entire thing and uh paste it let me copy this go back to CLA writing example paste from Tex paste it create a style and the moment you paste it CLA starts understanding what is content that is part of the system prompt and it itself gives you a style title in my case it actually gave me contemplator analyst I'm not sure what is the new title it is going to come up with but yeah so contemplative analyst engage in deeply introspective methodical reasoning that prioritizes exploratory thinking and continuous self-examination once you enable this all you have to do is anytime you think that you need this kind of a contemplation or think thinking or reflection then you can go select default would be concise or normal you can go select contemplative analyst then you would have contemplative analyst with you so I'm going to give you two claw examples then I'm going to tell you what is the disadvantage of using this approach first of all there is a very simple question I get out on the top floor the third floor at street level how many stories is the building above the ground the right answer to this question is technically one story so the Prem is unclear it is asking me question around okay um have uh can you clarify the building's layout I said okay just go over with go with whatever you have got and then it says based on the information that the third floor is the top floor and is at the street level so there are zero stories above the ground level uh the entire building appears to be at or below street level and it does not give me the right answer because the right answer is there is one story above the ground level clot doesn't do it without the system prompt but once I enable the system prompt like you can see here it is concise but once I enable the system prop in this case it is the contemplative analyst then it says okay let me think through this carefully and then it says street level entrance is on the third floor which immediately presents an interesting spatial puzzle I need to carefully consider what this implies about the building structure if I'm getting out on the third floor that is at street level wait let me examine this means more precisely what what this means the third floor coinciding with the street level suggest that building must extend downward rather than this being a traditional ground up counting system so if the third floor is at the street level then logically there must be two floors below the street level this means no let me pause and reconsider and then it reconsiders everything and then it says the street level is equal to and it also ask the question am I being too HD okay let me verify my reasoning the street level is equal to third floor stories above ground is equal to floors higher than street level in this case the street level floor counts as one story above the ground no mention of additional floors and then it tells me exactly that there is only one um ground story about ground level which is the right answer this is all possible because the model actually contemplated it went through whatever the system prompt has been given like the one that we took from mahishi and then it gave you the right answer and probably I think this is one of the most uh Innovative and also exhaustive list of system prompt you don't have to exactly use the entire thing the disadvantage here is that you're going to use up a lot of tokens so if you have to if you happen to use this you are going to run out of tokens a lot and is this will fill up your context window so for a longer context model it is easier for you to use this for a shter context model it might not be wise for you to use it all the time and also you don't have to like if I if my prompt is is literally like can you write a joke about Elon Musk then you don't have to use this kind of system prompt to enable I think that's where CLA is like slightly easier I can always go and select here this is one of the most comprehensive system prop that I've seen and there are a lot of different parts that you can use tailor it for your own requirement whatever that you want to do with and then you can change it accordingly I link the tweet and also the system prompt in the YouTube description if you have got your own interesting system prompt let me know in the comment section but otherwise this is a pretty useful system prompt for you to use if you do not care or if you are not concerned about context window a lot thanks to marishi for sharing the system prompt and also see you in another video Happy