This SIMPLE Trick will Change LLM Accuracy !!!

[@bonadio60]

This is really good, this shows that even simpler models with a good prompt can do very well. Thanks for sharing. Congrats to Maharshi-Pandya

[@aaronabuusama]

honestly this feels like a new model! its made claude so much more useful

[@BrianMosleyUK]

Yes, I've been using this today - it's terrific.

[@TechFrontiers-eg6wz]

one of the most useful videos I have watched in a long time. Prompt works like magic, well not quite but it works very well

[@MikaelHindsberg]

That’s really cool and useful, thank you.

[@1littlecoder]

Thanks the og author for sharing it!

[@MikaelHindsberg]

⁠​⁠will do. And an especial thanks to you for your video “I wish every AI Engineer would watch this” from a few months back. It was hugely helpful when I was just getting started with agents.

[@HashtagTiluda]

Seems like DeepSeek is using an in-built system prompt which is similar to this. It just got the answer in one go.

[@1littlecoder]

The thinking model ?

[@HashtagTiluda]

@@1littlecoder No, the regular one.

[@1littlecoder]

@@HashtagTiluda that's very interesting

[@HashtagTiluda]

@@1littlecoder Yes, I watched your video and then I thought why not use this problem on DeepSeek and see if it can solve without this system prompt. To my utter surprise it did solve it. DeepSeek is really amazing!

[@alx8439]

Give it a try with smaller open weights models

[@jmirodg7094]

Thanks that is useful!

[@xXWillyxWonkaXx]

But this defies the purpose of using tree of though then no? Plus, isnt that where Zod and Pydantic comes in?

[@MM-24]

what about a real use case for this? Im not using AI to solve basic riddles ...

[@ZacMagee]

This is the problem. Sure, some of it can help but there are 2 issues. This is not a system prompt. This is simply adding to the system prompt to help guide the model. The second issue is all its doing is clogging up the context window. You can get the same thing from simply adding "think through this step by step, never make things up."

[@thebigbigdaddy]

How is this in performance difference when using Professor Synapse text? Thoughts?

[@xXWillyxWonkaXx]

It's the same thing, almost. It simple just tells the LLM to accurately deliver knowledge-rich answers rather than extended reasoning loops.

[@maxuscz]

Prompt good working only in english language. Im try change it lie - reply in my language or give me reply in language XXX and it stop to working.

[@superdrew907]

The prompt has 2 many characters the response has 1679/1500 . . can anyone help me out here what am I doing differently?

[@1littlecoder]

As i mentioned on the video for chatgpt, cut the last part

[@superdrew907]

@@1littlecoder I cut it in half. just like the video, from output format down. Im still over @ 1694/1500. am I cutting to early?

[@cagycee5296]

Nvm. I got it.