It started making broadcasts instead. 134 models in one tab, 16 agents, and the whole pipeline.
A year ago I could generate a monkey. I could not generate a broadcast.
That sentence is the entire reason this workflow exists, and it took me about four months to work out what it actually meant.
The monkey was never the hard part. Fur under arena lights got solved a while ago. What kept giving my clips away was everything the monkey was standing inside: the scorebug that drifted between cuts, the sponsor boards that moved when nobody moved the camera, the cage mesh that passed behind a fighter instead of in front of him.
Your brain reads that layer before it decides anything. Get it right and the clip gets filed under "saw it on TV." Get it wrong and it gets filed under "AI slop," no matter how good the subject looks.
So I stopped shopping for a better video model and started building a pipeline around the layer that actually fails.
Last month that pipeline shipped 31 clips and brought in $14,300. One of those clips did 359,000 likes, against 46,000 for real broadcast footage of an actual title fight posted the same week.
Sits directly under the dek. Both numbers are real, taken from the source posts
Here is the whole thing
I Was Using AI Backwards
My old loop was: open a model, describe something cool, generate, look at it, generate again.
Nine tabs. An image subscription, a video subscription, an upscaler, a voice tool, a music licence, a captioning app, a background remover, a folder of prompts that had stopped making sense, and a spreadsheet nobody updated.
The output was fine. The problem was that every clip was a fresh gamble. I was using an enormous amount of speed to manufacture untested ideas faster.
The fix was not a better prompt. It was giving each stage one job and never letting a stage do somebody else's.
Map of the five stages, placed before the reader walks through them
Step 1 · Find the Format, Not the Idea
I do not ask what is trending. Raw view count tells you almost nothing: 800,000 views is unremarkable if the account normally does two million.
What I want is the opposite. An account that usually does 20,000 suddenly posting something at 600,000. That is an outlier, and something inside it behaved differently.
Then I throw away the topic entirely. If the winning clip is about a fruit being cut open, "fruit videos work" is the useless reading. The useful one is:
familiar object appears immediately → one detail looks wrong → viewer waits for it to open → each reveal escalates → final frame explains what they were looking at
That skeleton works on a product, a character, a fight, a kitchen, a machine. I am not copying a video, I am extracting an attention mechanic.
Step 2 · Hire the Director
This is where research becomes production, and it is where Picsart enters.
Picsart put 16 agents live on May 21. They are not chat assistants, they are briefs with a job title: Marc runs end-to-end video, Reeva directs reels, Mila enforces style consistency across a set, Nora handles localisation, Sam runs creative operations.
The one I use is Lina, the cinematic short-film director. You hand her the format plus your subject, and she comes back with the thing I used to write by hand at two in the morning: a scene-by-scene plan, pacing built around the music, shot order, and where the payoff lands.
The difference is not subtle. Asking a model for an idea gets you a suggestion. Giving an agent a proven format and a subject gets you a plan with a reason behind every cut.
Step 3 · Race the Models
The AI Playground is the part I underrated for months.
It carries more than 134 models now, and the recent additions are the ones that matter for this work: WAN 2.7, Seedance 2, Kling 3 Omni, Recraft V4, Nano Banana 2, and Lyria 3 for music.
The thing that changed my week is not the count. It is that the same prompt runs across different models from one box, against one credit pool, and I pick the winner by looking rather than by reading benchmark threads.
For the broadcast work, Seedance 2.5 is what I lean on. It takes up to 50 references, images, video and audio together, and locks a character's face, wardrobe, voice and world across a whole sequence. That is exactly the failure I described at the top. It also does 20 languages with matched lip-sync, which turns one production into a dozen markets.
And it is all reachable the way you actually work: web, the desktop app that shipped in August, or wired into your own pipeline over REST, CLI, MCP and SDK.
One Visual Rule Fixes Most of It
Before any full build, I write down one rule and refuse to break it.
One lens. One grade. One character description, copied verbatim into every shot.
It is boring, and it is the single highest-leverage thing in this document. AI video falls apart when each scene gets treated as its own generation: shot one looks great, shot two looks great, and together they look like three different films. Consistency beats individual shot quality almost every time.
Step 4 · Build the Broadcast, Not the Subject
Here is the part I would tell anyone starting today.
Spend your effort on the frame around the thing, not the thing.
For the fight clips that means: a scorebug that counts down continuously instead of resetting between cuts, sponsor branding sitting where it always sits, the cage passing in front of the fighters and occluding them correctly, a cut pattern that goes wide, low, mat level. Broadcast has a grammar, and the grammar is what your audience actually recognises.
The subject is the cheap half now. The context is the expensive half, and almost nobody is working on it.
Step 5 · Publish and Measure
Publishing is not the end of the loop, it is the research stage for the next one.
Every clip gets tagged by format, and the number I watch first is completion, not views. If my normal clip finishes around 30% and one structure repeatedly clears 40%, that tells me more than a single spike ever will.
That format goes back into stage one, and the next batch starts with more information than the last.
The Part That Made $14,300 Possible
Opens the monetisation section. Doubles as the image for the announcement post
The money did not come from one viral hit, and it did not come from finding a secret model. It came from turning content into a repeatable line, because a line can be sold three different ways at once.
Platform monetisation pays when distribution and retention are both there, which is why completion rate is the number I track first rather than views.
Done-for-you clips are the same pipeline pointed at someone else's brand. A client who needs short-form creative without a production company is buying the process, not the render, and the process is already built.
Placement is the third. Once a clip reliably holds attention, a relevant product can sit inside it without breaking anything, and that is worth more than a banner because nobody skips it.
One production loop, three ways to bill it. That is a far more interesting position than trying to make one perfect video.
What This Actually Replaced
Sits inside this section, before the paragraph that reads the numbers out
Nine tabs became one, and the honest accounting looks like this:
134 models against a single credit pool. 16 agents I brief instead of nine tools I operate. 4K coming out of the same box that made the clip. 20 languages with lip-sync instead of a translator and a re-record. A desktop app since August, and an API when I want it in my own pipeline.
I am not going to pretend the closed frontier models never win a shot. They do. But the gap that justifies nine subscriptions did not survive contact with a playground where I can race them.
What I Stopped Doing
I stopped asking for original ideas every morning. I stopped choosing a direction because one generation looked cool. I stopped treating each upload as an unrelated experiment. I stopped changing my visual style weekly.
Mostly I stopped confusing originality with starting from zero. Short-form attention already runs on known mechanics: transformation, escalation, countdown, comparison, reversal. The creative work is what you put inside them.
The Bottom Line
Most AI content starts with a blank prompt. Mine starts with evidence.
Find a format that already beat its own channel. Strip the topic. Hand the skeleton and my subject to a director agent. Race the models on the shot. Lock one visual rule. Build the broadcast around the subject, not just the subject. Publish, measure completion, and let that decide what gets made next.
Nobody fought in that octagon. 359,000 people watched anyway.
That number is not about the monkey. It is about everything the monkey was standing inside.
$14,300 last month, and I would rather run the loop again than wake up tomorrow and ask AI for another original idea.
Lina, AI Film Producer and Director Agent by @Picsart
Link: picsart.com/ai-agents
