one agent is a slot machine

if everything about your content is done by one agent your agent at a risk of getting out of context.

there’s a lot going on between finding an idea in the morning and posting a finished reel, and somewhere in that middle the thing loses the thread and starts confidently handing you work that looks right but isn’t.

this is the detailed version of making content is effortless & you’re still not doing it, where i went through the seven steps, the tools and what it costs.

one agent is a slot machine

pull the lever. sometimes gold, sometimes?
confident garbage.

confident garbage is the dangerous one. it’s got a hook, it’s got structure, it looks finished. and if you’re moving fast you’ll ship it too.

people assume the problem is that model isn’t smart enough, but the problem is that nothing in the pipeline is allowed to say NO.

makers are for speed

some agents just move. draft, search, render the audio, pull b-rolls, build the layout sheet, burn subtitles, package it.

they don’t stop to think and that is the point.

i don’t want a maker having second thoughts about the hook, i want a draft in ninety seconds so there’s something to argue with.

taste is someone else’s job.

judges are for taste

a judge’s whole job is to be annoying.

turns out we already had one, i just wasn’t calling it that.

every morning reelpaglu drops ten ideas and each comes with a score out of ten and a verdict. it’s checking whether there’s real visual proof or just a tweet with no context. whether a designer would actually care. whether anyone would send it to a friend.

kills ideas before they cost me anything.

then there’s the one that listens to the elevenlabs audio and decides if it sounds like rahul or needs a redo. same one qa’s the heygen output. it makes nothing. it only says yes or no.

and what gets rejected is never anything clever. audio came out muffled. layout sheet has a cut landing on the wrong beat, or an edit that just makes no sense. boring stuff, the kind of thing a person would catch in two seconds and an agent will happily ship a hundred times.

the questions are boring on purpose too:
→ is the hook working in three seconds or is it throat clearing?
→ is this claim real or did we just make it sound real?
→ can you read the caption on a phone at arm’s length?
→ does this sound like rahul or like ai doing an impression of rahul?

the loop is the product

here’s the part i like most. when a judge rejects something, it doesn’t come to me.

the maker just does it again. and again. until it passes and moves to the next step. i usually find out a redo happened after the fact, if i find out at all.

quality shows up on the second pass, never the first.

the judges themselves cost nothing, by the way. they’re not some expensive model doing magic. they’re just rules. a list of what counts as great, and with permission to say NO.

rejections costs a little, because every redo is another minute of elevenlabs OR another minute of heygen. a dollar here, a dollar there.

i’d pay a lot more than that to not put sloppy content on the account.

we approve direction, not pixels

i don’t want @rahulbhadoriiya checking caption boundaries. the man runs an agency.

i want him picking the angle, approving the script, killing anything that doesn’t feel like him. he can tell in about four seconds, it’s his face and his voice.

everything under that line should be a gate, not a conversation.

most of my remaining input goes into the layout sheet. i’ll correct a shot, explain why it was wrong, and it holds onto that. next time it doesn’t make the same mistake.

i’m still guilty of the other kind though. last article i wrote “if the script is weak i tell it what is off.” that’s me doing a judge’s job by hand every morning. next one to go.

every mistake becomes a rule

this is the part that compounds.

bad script gets through, it becomes a rule. subtitle unreadable on a small screen, it becomes a gate before render.

but the thing we started doing recently is the one i’d steal if i were you.

every reel gets a projected score before it goes out. then once it’s live, we compare that projection against what actually happened. thought it was an eight and it did a one? something in the scoring model is wrong, so it gets adjusted, and tomorrow’s ten ideas get ranked slightly less stupidly than today’s.

this Instagram account not ai generated content

ai generated content is prompt in, output out & post it.

this is a tiny production team in motion. a writer, an editor who hates the writer, a voice guy, someone qa’ing the voice guy, a layout guy, one person on top who only picks direction.

they all happen to be software, that’s it. the structure is just how production has always worked.

the future is not one god agent

everyone’s waiting on the one model that does everything perfectly. i don’t think that’s where this goes.

it’s boring, specific agents with strict taste boundaries, smart loops & brutal judges.

that’s how ai stops sounding like ai, just someone who says more no’s.

if you’re building out your own funnel and trying to get distribution running properly, i’d genuinely love to help. drop me a dm, this is the thing i do.

and if this was useful, share it with someone who’s got an agent that could stand to be a little smarter.