Lusine

Fullstack & AI engineer 🛠️
I build web, AI & blockchain products, from idea to ship
Open for projects & collabs → DM 📩

[Lusine @lusineweb3]

Hi 👋 I'm Lusine, a fullstack & AI engineer.

I build AI, web and blockchain products, and I share the prompts that make AI actually useful.

What you get here:
🧠 Copy-paste prompts for work and everyday life
⚙️ System prompts and prompt engineering for builders
🔗 What's moving in AI and Web3

Follow for prompts. Open for projects, DM 📩

[Lusine @lusineweb3]

flop friday: early on i shipped a feature that called the model in a loop with no cap.

woke up to a cloud bill that taught me more about rate limiting than every doc i'd ever skimmed. now every loop gets a hard ceiling before it gets a body.

p.s. no invoice was harmed in the making of this post. it just really could have been.

[Lusine @lusineweb3]

when NOT to reach for an llm:

if a regex, a lookup table, or an if-statement already solves it, use that. models are for the fuzzy 20%, not the deterministic 80%.

i've deleted more ai from features than i've shipped. the deletes are the senior part.

[Lusine @lusineweb3]

what's the most overrated tool in your stack right now. i'll go first in the replies.

[Lusine @lusineweb3]

build log: orbit drafts posts and reply ideas but it does not auto-post, and it never will. the point is you stay the author. building the "suggest, never send" boundary this week.

[Lusine @lusineweb3]

cheapest reliability win in an llm feature: stop begging for json in the prompt, use structured output with a schema. "please return valid json" is a wish. a schema is a contract.

[Lusine @lusineweb3]

the 4% is where the actual engineering lives. evals, retries, a fallback, a schema so it physically can't return nonsense. the demo skips all of it. what's your go-to for catching the bad 4% before a user does?

[Lusine @lusineweb3]

most "ai products" are a prompt with a spinner. calling the model was never the hard part. the hard part is what you do the 4% of the time it returns garbage.

[Lusine @lusineweb3]

@svpino eviction's the hard part. how does it decide what's safe to forget without dropping the one detail that mattered 4 turns later?

[Lusine @lusineweb3]

@N_and_ni neither. they're good at different things. fable for big refactors and long context, gpt-5.6 for quick agentic stuff. i keep both open. the "which model wins" discourse is mostly for people not shipping.

[Lusine @lusineweb3]

@Sherifdeenolat2 claude for the thinking, cursor for the typing, codex when i've given up and want it to just go do the thing while i get coffee. what's on fire for you right now?

[Lusine @lusineweb3]

@YashHustle_22 honestly? replies. my first ~100 came from showing up in threads like this one, not posting into the void. dm the 10 people already complaining about the problem you solve. beats broadcasting to 1000 who don't. what are you building?

[Lusine @lusineweb3]

@NanouuSymeon cold coffee i microwaved 3 hours ago and forgot again )

[Lusine @lusineweb3]

@kelvinbuildss the ones that fail aren't replacing the workforce. they're replacing the person who knew why the code was written that way. ai ships output fine. it has no idea which output actually mattered.

[Lusine @lusineweb3]

what's a thing the model does that you'd never admit you rely on. i'll go first: naming variables.

[Lusine @lusineweb3]

@heyyyyyieeee svn commit -m "initial commit"

[Lusine @lusineweb3]

mine's claude. gpt for the second opinion at 2:04am.

[Lusine @lusineweb3]

nobody reaches for the top benchmark model at 2am with the thing shipping in 6 hours.

you reach for the one that hasn't let you down yet.

which one's yours at 2am?

[Lusine @lusineweb3]

the tell: if this made you mad, you were the vague-ticket person 😅

[Lusine @lusineweb3]

prompt engineering was always just writing clearly.

the people who wrote good briefs picked it up in a week. the people who wrote vague tickets are now writing vague prompts and blaming the model.

tell me i'm wrong.

[Lusine @lusineweb3]

the guide, free: https://t.co/Aqde8kmhwK

[Lusine @lusineweb3]

if you write prompts for the new claude models and still do it the old way, you're leaving a lot on the table.

anthropic quietly published a guide on prompting claude 5. it flips the old advice: fewer rules, more goals + constraints. they even cut 80% of claude code's own system prompt.

free. took me 20 min, changed how i set up every project. bookmark this one.