The most valuable queries are the ones I know their answer in advance. It’s just that I am too lazy to craft the answer myself. Just like you did. If I were assigned to do your exact same task with terraform (something I don’t have much experience with), I wouldn’t be able to successfully query the LLM to do the job.
Ask HN: What's your most valuable query to an LLM?
11–15 of 15 posts
Re: Ask HN: What's your most valuable query to an LLM?
#12Earlier quoted context omitted.
another example just from today: I merely selected entire json file, two word prompt: "generate schema" one-shotted 600 lines json schema, unsolicited 200 lines typescript schema, 150 lines python dataclass model and a README!!! completely unsolicited! (cursor agent mode)
Have you checked it yet? Last time I did that from an Open API file rather than transform it, as I was lazy, it hallucinated a bunch of properties and left a load off too. That was a year or so ago.
a lot changed in a year
Re: Ask HN: What's your most valuable query to an LLM?
#13My useful local stack: Ollama + Llama 3.1 8B (chat/RAG), VSCode + Continue.dev for coding, Qdrant for lightweight retrieval, 64 GB RAM desktop. Works offline with low latency; main gotcha is context blow-ups → chunking + token monitoring.
Bonus: I A/B prompts/styles in the cloud first, then reproduce in local ComfyUI. Disclosure: I built flux1krea.app to baseline prompts/styles before moving local: https://www.flux1krea.app
Re: Ask HN: What's your most valuable query to an LLM?
#14I like to digitize my handwritten notes. For the basic transcription, there's isn't much prompting to do, but I do tell it some stylistic choices I make and how to interpret those as markdown, like open to-dos as circles and closed to-dos as circles with Xs to their markdown equivalents. I also upload some common templates for things like my weekly reviews and tell it to use a template if applicable. I'm sure if I dr…
I'm curious how clean is your handwriting? I just did a test run with Claude (pro plan) and it could not understand my iPad + iPencil notes. I expected this since they are close to chicken scratch. ChatGPT told me "I ran OCR on your notes, but the handwriting/scan quality makes the transcription pretty rough—lots of words are garbled."
Then I tried a handwritten note that was some of my average handwriting and it did okay-ish. I think it would be acceptable to use if I also store the original note image with the transcribed text and for basic searching.
I agree with you I prefer writing physically but need a system for digitizing.
Re: Ask HN: What's your most valuable query to an LLM?
#15I like to digitize my handwritten notes. For the basic transcription, there's isn't much prompting to do, but I do tell it some stylistic choices I make and how to interpret those as markdown, like open to-dos as circles and closed to-dos as circles with Xs to their markdown equivalents. I also upload some common templates for things like my weekly reviews and tell it to use a template if applicable. I'm sure if I dr…
I'm curious how clean is your handwriting? I just did a test run with Claude (pro plan) and it could not understand my iPad + iPencil notes. I expected this since they are close to chicken scratch. ChatGPT told me "I ran OCR on your notes, but the handwriting/scan quality makes the transcription pretty rough—lots of words are garbled." Then I tried a handwritten note that was some of my average handwriting and it did…
I'm honestly not sure how my handwriting is - I haven't had someone ready my handwriting anything outside of math or computer science in a decade. I would guess my handwriting is an average man's. Legible, but might take a sec to read at points.
If I have lines that run to the end of the page and stuff gets garbled, Claude does struggle with that and says it added some assumptions in.