How Prompt Compression Can Make You a Better Prompt Engineer

[@Juan_SZ]

I recently found your channel, and with each video I watch, I learn quality content, very different from what others offer. I am not saying others are worse, but your focus is a mix of more technical explanation while still keeping it basic for non-ML experts. This is very valuable content; thank you.

[@Mark_Kashef]

Juan, it means a lot to hear that! Thanks for the positive encouragement, I'll keep it coming 🦾

[@jalengonel]

Lyrics and songwriting have started to have greater significance to me since I started prompting. Songs are basically encapsulated stories whose words say 100x more than the sum of their parts. It’s not always the details, more so it’s about evoking the “energy”. Listening to Billy Joel has been a really great exercise for communciating with brevity.

[@Mark_Kashef]

you may have given birth to a new way of prompting :) "You are a sophisticated prompt engineer, you write your prompts similar to songwriters writing songs. High value per word written, with compressed instructions only keeping the completely necessary commands" riffing of course haha

[@mwdcodeninja]

I really appreciate the technical breakdown alongside the llm approach.

[@Mark_Kashef]

pumped to hear that! Appreciate the feedback 🦾

[@pengouin]

One of the best youtuber about ai for sure, thank you.

[@Mark_Kashef]

thanks so much! pleasure to hear that you think that haha (glad one of us does)

[@tonyblack2141]

I see a notification from you... I watch, cos there's always more than a nugget to get. Thanks mate.

[@Mark_Kashef]

I see a 'new comment' notification, I open up the YT app to respond to your kind feedback haha!

[@tonyblack2141]

😂... much appreciated

[@MJFUYT]

Excellent content Mark. You are my go-to for everything ai prompting. Thanks for sharing your knowledge.

[@Mark_Kashef]

thank you so much! Love hearing that. Glad I can provide value to you.

[@EdwardAustin]

Dude, fantastic video.

[@Mark_Kashef]

thanks so much Edward! Glad you liked it 🦾

[@davidtowers7851]

Mark, as usual, Extremely useful and to the point.

[@Mark_Kashef]

Appreciate you David! Thanks for the positive feedback loop on this

[@stefmedia]

Cool stuff! You always have interesting stuff out of the ordinary and very relevant. Well done. Thank you.

[@Mark_Kashef]

Thanks Stef! I'm glad my innate mad scientist is valuable to you, appreciate the feedback 🦾

[@goodspeed199]

This is so good. Thanks a lot Mark for sharing this with us.

[@Mark_Kashef]

pumped you enjoyed it! wasn't sure how this one would be received, so I appreciate the feedback loop ➰🦾

[@goodspeed199]

​@@Mark_Kashef Yes, I would love to see more of these topics but these diversity of your content are really special. I haven't had time yet, but now I'm heading over to Gumroad. 😀 Once again thank you very much Mark.

[@janalgos]

Puregold as always

[@Mark_Kashef]

thank you! Appreciate you 🙏

[@apricotcomputers3943]

I'm now a believer! 😂

[@bgmspot7242]

Could you please do course or video on prompt engineering (A to Z). I want to learn from you 🎉🎉🎉

[@Mark_Kashef]

I have a truly special video coming, but it has to wait for a few more weeks -- stay tuned on this request 👀

[@mzafarr]

Love your videos man! I have been working on a project for quite some time, but failed to get it to work perfectly, I'm trying to build an AI system that can solve legal cases (I have a db of laws, I retrieve the relevant laws using some techniques) now the issue I am facing is that none of the LLMs are great at solving law cases, I tried making tweaks with prompting and so many different techniques, but for complex cases it just isn't good, o1 is much much better than all of the rest I tried (claude, gemini, 4o, llama 3.2) but it also doesn't get correct all the time, I think main issue is that these LLMs might not be trained on solving legal cases at all it's maybe just using basic common sense for interpreting the laws I provide to it. Any help/tips/advice could be of EXTREME BENEFIT for me. And if you need more in detail information about what exactly my current system is doing, I can explain that all. Thank you, Mark.

[@Mark_Kashef]

Thanks so much, means a lot to hear that! I read this a few times to mull over it, and to me the LLM you choose is important, but not 'as' important as your method of: 1) retrieval (vector database selection - supabase, chroma, qdrant, pinecone, or something like assistants api vector db, or pinecone assistant) The way you're storing these laws -- are they raw documents 'with' the laws, or have you distilled the laws into a more palatable, lower noise format like FAQ format, list format, thematic categorization of these laws in markdown etc. 2) The best performing LLM we've seen working with law firms is Claude Sonnet 3.5; I also think their new pdf API can be a helpful weapon to use for this project in some capacity I would take a look at the company Spellbook AI (no affiliation) to see what they're upto; they're tackling your problem in an interesting way, you may draw some inspiration from it

[@xaamirx]

I'm in Qatar. Let me host you for a coffee 🙏

[@markvandermeulen1612]

Mark the video is 7 minutes too long. Are you that friend?

[@Mark_Kashef]

Guess I am 🤷🏽‍♂️