field notes
// field notes / teardown

jazzelle@thehotpath ~/field-notes $ cat the-carousel-engine.mdx

one command, ten seconds, five dollars a year

one command, ten seconds, five dollars a year

2026-05-30 · 6 min

every carousel on @thehotpath comes out of a pipeline i built, and i have never once opened canva.

the short version: it researches what's actually trending in ai, hands me three prompt directions, and once i pick one it renders a full terminal-aesthetic carousel in about ten seconds. it runs on under five dollars of tokens a year. the long version is the part worth reading, because the cheap fast number is not the point. the point is where i kept the human.

// the long way around, even when it's fast.

why not just use canva

because canva makes you look like everyone on canva. the templates are good, which is the problem; good templates converge on a single average look, and the average is invisible. i didn't want a tool that made design decisions for me. i wanted a tool that made my design decisions faster.

so i built one. node and playwright doing the rendering, my own tokens, my own type, no drag and drop. the aesthetic is the terminal because that's mine, and a pipeline can hold a point of view in a way a template library can't.

how it actually runs

it goes in four moves.

it researches. before it proposes anything, it looks at what's trending in ai right now, so the post is timely instead of evergreen-but-late.

it offers three. not one. three prompt directions, each a different angle on the topic. this is the load-bearing step and i'll come back to it.

i choose. i pick the direction that's actually mine. the machine does not pick. this is non-negotiable.

it renders. node and playwright turn the chosen direction into a full carousel, on-brand, in about ten seconds. then it's a post.

the part that matters: it offers, i decide

i could have built this to post on its own. full automation. wake up, check the feed, the bot already shipped three carousels overnight. it would have been easy and it would have been wrong.

here's the mechanism: because i choose the prompt every time, the system curates toward my taste instead of away from it. each choice is a small vote. over weeks those votes compound, and the pipeline gets sharper at proposing things i'd actually pick, because it's learning the one thing it can't generate on its own, which is judgment. an autonomous version would drift toward the average and i'd never notice until the whole feed looked like nobody. the human in the loop is not a safety feature here. it's the entire quality engine.

the receipt

ten seconds a post. under five dollars of tokens for a year of them. no subscription, no template, no canva. and a feed that looks like one person made it, because one person did, just faster.

// the computer does the fast part. i keep the taste. that's the trade i'll take every time.

$

ask the agent yourself