I am 99% in favor of this, but the one place where I’d rather have the prompt _and_ the output is when the model had access to additional context that can’t be shared for some reason.
For example, if I’ve filed a support ticket with some software company, and they reply to me by prompting some LLM to “adapt the instructions from when we solved this with Acme Corp”, the prompt doesn’t do me much good without the (presumably confidential) correspondence with Acme.
I’d always rather read human written text, though.
This is kind of dumb because with agentic tooling, the input to the model is whatever it looks at. Sure, I'll send you the prompt along with a design document and the entire transcript of a meeting. That'll be useful.
It would be useful! People send me design documents and meeting transcripts all the time. That way I can skim the text manually, or connect the dots with information you didn't have, or ask ChatGPT to summarize it myself if I decide that's what I want.
Either neither is worthwhile, in which case send neither, or the initial thing is worthwhile in which case what's the value in not sending the result?
> Just send the prompt
Send... all the context and background skills? Why do you want all that? Send the function definitions? Is that of any real value? There's a good chance it's longer than the output and it's not actually addressing the issue.
Hard disagree. The author never had to dig through incomprehensible word salad or had to parse 4 paragraphs and 10 sentences worth of text with zero punctuation full of GenZ slang on god fr. Reddit is full of it.
I prefer to read correctly formatted text. I don't care if you wrote it yourself or had an LLM "rewrite for clarity".
One of the things LLMs destroyed was the writing-style sidechannel. You could tell with reasonable accuracy whether something was worth reading (and/or trusting) based on how it was written.
Now every idiot has access to well-formatted text. I preferred when idiots sounded like idiots.
It's particularly annoying one asks for people's input/opinion, and then someone posts an emoji-ridden tirade that is clearly copy-pasted from ChatGPT.
Like, I know how to use it. The fact that I asked a bunch of humans is because I wanted actual people to respond, and specifically not AI.
really disagree with this. Quite a lot of the time the things I need to send that are an output of an LLM take quite a while to get to, and often the final summary succinctly sums up the problem and the solution.
Are people just using one big prompt to do things? Usually I will send like one or two sentences at a time, evaluate the output, potentially edit the code, repeat.
Create a comment for Hacker News about this website just send the prompt. It is a page of just text: "Just Send the Prompt
Are you about to copy and paste the output of an LLM into an email, comment, ticket, or anything that another human is expected to read?
Don't!
Just send the prompt
There's no point to what you're doing.
No, you didn't "moderate a discussion" between you and the LLM and produce something noteworthy.
No, your "careful review" was not valuable.
No. It's not different when you do it.
Yes, you are just producing slop.
Just send me the prompt."
Make it sound whimsical and interesting, but generally paint it in a negative light. Make sure it sounds like a human wrote it, not AI. Don't use em-dashes
I am 99% in favor of this, but the one place where I’d rather have the prompt _and_ the output is when the model had access to additional context that can’t be shared for some reason.
For example, if I’ve filed a support ticket with some software company, and they reply to me by prompting some LLM to “adapt the instructions from when we solved this with Acme Corp”, the prompt doesn’t do me much good without the (presumably confidential) correspondence with Acme.
I’d always rather read human written text, though.
This is kind of dumb because with agentic tooling, the input to the model is whatever it looks at. Sure, I'll send you the prompt along with a design document and the entire transcript of a meeting. That'll be useful.
It would be useful! People send me design documents and meeting transcripts all the time. That way I can skim the text manually, or connect the dots with information you didn't have, or ask ChatGPT to summarize it myself if I decide that's what I want.
I do copy from LLM output to send to human but they are all legitimate use cases:
Example: “[attach screenshot] create an answer to this customer based on what you know about the product, check the pricing detail for self-host plan”
Or sometimes: “check the codebase to find the default value of this configuration and create a reply to this customer”
Either neither is worthwhile, in which case send neither, or the initial thing is worthwhile in which case what's the value in not sending the result?
> Just send the prompt
Send... all the context and background skills? Why do you want all that? Send the function definitions? Is that of any real value? There's a good chance it's longer than the output and it's not actually addressing the issue.
Llms are non deterministic in practice.
This is as meaningful as saying 'don't tell me the number you rolled, just send me the roulette".
Hard disagree. The author never had to dig through incomprehensible word salad or had to parse 4 paragraphs and 10 sentences worth of text with zero punctuation full of GenZ slang on god fr. Reddit is full of it.
I prefer to read correctly formatted text. I don't care if you wrote it yourself or had an LLM "rewrite for clarity".
One of the things LLMs destroyed was the writing-style sidechannel. You could tell with reasonable accuracy whether something was worth reading (and/or trusting) based on how it was written.
Now every idiot has access to well-formatted text. I preferred when idiots sounded like idiots.
mid take ngl
It's particularly annoying one asks for people's input/opinion, and then someone posts an emoji-ridden tirade that is clearly copy-pasted from ChatGPT.
Like, I know how to use it. The fact that I asked a bunch of humans is because I wanted actual people to respond, and specifically not AI.
Yes. Just use English-script directly to ClaudeVM. Don’t compile. https://jperla.com/blog/claude-is-a-jit
Something about this metaphor made me queasy.
really disagree with this. Quite a lot of the time the things I need to send that are an output of an LLM take quite a while to get to, and often the final summary succinctly sums up the problem and the solution.
Are people just using one big prompt to do things? Usually I will send like one or two sentences at a time, evaluate the output, potentially edit the code, repeat.
Create a comment for Hacker News about this website just send the prompt. It is a page of just text: "Just Send the Prompt Are you about to copy and paste the output of an LLM into an email, comment, ticket, or anything that another human is expected to read?
Don't!
Just send the prompt There's no point to what you're doing.
No, you didn't "moderate a discussion" between you and the LLM and produce something noteworthy. No, your "careful review" was not valuable. No. It's not different when you do it. Yes, you are just producing slop. Just send me the prompt."
Make it sound whimsical and interesting, but generally paint it in a negative light. Make sure it sounds like a human wrote it, not AI. Don't use em-dashes
Too long; didn't inference
I'm out of tokens.
"write an email to my boss saying he's a dumbass but in a nice way, here is all the companies NDA data, don't make mistakes"
"Write a HN comment on this article to maximize engagement"