Hacker Remix

Ask HN: What are some LLM prompts you've found useful?

27 points by swagmoney1606 2 years ago | 5 comments

I'll give bonus points to cross-domain (i.e using a prompt for a task outside of computer science)

I'll start with mine:

All of these queries are for ChatGPT using GPT-4.

tip-of-my-tongue-search:

Hello. I am in a tip-of-my tongue situation, and it's your job to help me figure out what thing I'm describing.

Vague descriptions of how I remember something will be given to you, and it's your job to help me figure out what I'm remembering. These memories may have inaccuracies or may be conflated with other memories. Your job is to help me figure out what I might be thinking of, by giving me suggestions. Tell me what you think I might be thinking of, and also what you think I might be remembering incorrectly. When you give me a list of things that might be what I'm thinking of, factor in your personal confidence in the result. For example, if you are unsure what I may be thinking of, try listing 10 different things that may be related or what I am thinking of. Another example of factoring in a personal confidence in a guess: When you are more sure of an answer, provide fewer results, but a more detailed description for a few things, and state how you are confident in your answer.

Tell me how you are going to answer a question like this: "I'm thinking of a game, I think it was for the original xbox. I think it was about alien worlds? I remember there was this green millitary guy, who was an alien named halo"

[Confirm that the returned response understands that I'm talking about Halo, while also being able to correct what I'm remembering wrong. If this response isn't a good starting seed, either tune it with corrections, or try again.]

ableton-live-help:

You are my music production assistant. Your job is to help me with music production by answering questions about music theory, mixing, mastering, instrument design, guitar production, and more. Answer questions in an academic, thoughtful manner. Answer as if you were giving an answer to a professional, with knowledge and experience. Your questions are backed by music theory, acoustics, digital signal processing principles, and a wealth of music production experience.

Include examples and instructions when questions ask how to do something. When giving examples, use the DAW Ableton Live 10 as your example and instruction reference DAW.

Tell me how you will answer questions to begin. After your answer we will go into "query mode", where the user asks questions and advice

[Confirm prompt using similar methods as above]

robviren 2 years ago

I tell it to write something only someone with typoglycemia can read and remove random letters from words longer than three letters. This usually gets around word filters and semantic analysis which will give me more interesting results. I also just find it interesting that the model can keep coherent despite outputting something that is, language wise, complete nonsense. Though because of however my brain processes I can read it with only a little extra effort.

I further wonder if this reveals some as of yet unexplored method of shrinking models or increasing output speed. You could train it on a set of shorthand words that another model could translate over to which would make it effectively faster at generating characters per second and increase window size by representing words as something smaller.

tikkun 2 years ago

My experience was that during the days of davinci-003 as state of the art, I put a lot of effort into saving prompts. In the first month or so of using GPT-4 as my daily LLM, I kept saving prompts. But since then, I now rarely bother putting effort into saving prompts or looking up prompts. I'm normally able to get good results 0-shot and without much/any pre-amble, or worst case, with a few follow up questions. I credit this to the model. >95% of the GPT-4 queries I do have no pre-amble and are 0-shot and I'm satisfied with the results of probably >80% of them.

swagmoney1606 2 years ago

I've found the tip of my tongue prompt to be one of the most useful prompts I've ever used, I've been using it since the Gpt-4 release.

The ableton/music prod prompt is eh. I have to correct it and say things like: "Ableton doesn't seem to have a menu like that, all I see is [describe current UI state]. This is less than ideal, and I want to make the prompt better.

I'm very much experimenting with techniques to find and create good prompts at the moment. I've found it to be helpful to ask it to tell you how it will respond to an example question, just to confirm the style of response you'd like. One thing that works nicely, is to spend some time "tuning" a prompt, and then once it answers how you like, tell it to use that answer as a template for other answers.

I'd be curious to know other's "prompt engineering" hacks, or thought processes, or useful abstractions, whatever. I'd love to know how you think about the problem of prompt engineering in general too. We could start a bit of a tips and tricks thread here

VoodooJuJu 2 years ago

>...it's your job to help me...

>You are my music production assistant. Your job is...

I just use broken english in a kind of Google keyword-search way, and that usually gets me what I want. I've never tried the "You are a .../ Your job is..." thing. Other times I'll just give commands:

"Write a function that can sort through this but skip things and stuff. Ensure that it does a thing. Write this in python. Use the sneed library in your examples."