I Built an AI Agent That Makes Viral Videos

$14,700 last month. I approve the idea. The agent does the other four steps alone.

I spent 3 weeks building one agent. It decides what to make, makes it, ships it, then reads the result and adjusts itself.

I gave it one rule: find what is already winning before you build. Every video that hits leaves a public trail. The agent reads the trail, not the video.

Here's the full build.

The Pipeline

IDEAS → SCRIPTS → EDIT → PUBLISH → VIRAL VIEWS

find the winning format → write it for the cut → build the clip → post everywhere → read the result and feed it back

One agent holds all five. The last arrow points back to the first. That loop is the whole product.

The Stack

• Kimi K3: the brain that runs the loop and calls every other tool
• Seedance: turns a shot list into moving vertical clips
• ElevenLabs: the voice on every clip
• Blotato: posts to TikTok, Shorts, and Reels from one call

Four services. One agent orchestrating them. No human between the steps except the approve button on the idea.

Step 1. IDEAS: The Agent Finds the Hit Before It Builds It

80% of the result lives here. Every morning the agent scrapes 40 faceless channels in one lane and scores the day's breakouts:

outlier score = video views ÷ channel median views

A 20k-average channel drops one at 900k. Score of 45. Above 30 is a format you copy this week. It stores the winner's first frame and first line of text, strips topic from format, then runs one prompt:

1You are a short-form video strategist.
2Here are 20 videos that beat their channel by 10x, with hooks and outlier scores.
3
4[the agent pastes its 20 scraped rows]
5
61. Group them into 5 repeatable FORMATS (the shape, not the topic).
72. Take format #1, my subject is [niche], write 30 concepts in it.
83. For each: hook line (max 8 words) and the moment that carries the first 2 seconds.
9Rank 1-30 by scroll-stopping power. Cut anything generic.

I see 30 concepts on a proven skeleton, tap the three I like. The agent takes it from there.

Step 2. Formats That Keep Hitting

The agent logs every shape it sees break out and reuses them forever.

• "POV: you're the [role] when [event]": puts the viewer inside the scene in 1 line
• "What if [familiar thing] but [twist]": curiosity gap, no context needed
• The invisible-force gag: objects move on their own, a chaotic reveal at the end
• Oddly satisfying process: a start-to-finish build, no talking
• 1 vs escalation: 1 vs 100, level 1 vs level 100, small stake to huge stake
• The creature reveal: a normal scene, then the last 2 seconds flip it
• "Day in the life of [unexpected subject]": a rock, a medieval peasant, a deep-sea fish

Each shape carried a different subject to millions last month. The shape stays, the agent swaps the subject.

Step 3. SCRIPTS: The Agent Writes for Shots, Not Prose

Once I approve a concept, the agent writes for the cut:

1You are writing a 25-second vertical short in the format:
2"POV: you're the last lighthouse keeper on Earth."
3
4Write for shots, not paragraphs. Output:
5- on-screen hook text for the first 2 seconds (max 8 words)
6- 5 shots, one line each, described for AI video generation
7- one line of narration per shot, spoken pace
8- the final 2-second beat that makes people rewatch or comment
9
10Tone: quiet, eerie, no filler. End on an open loop, not a bow.

If the concept is thin, the agent flags it back instead of building a weak video. That flag saves more money than any tool in the stack.

Step 4. EDIT: Seedance and ElevenLabs Build the Clip

The agent sends the shot list to Seedance and the narration to ElevenLabs, then stitches a finished 9:16 vertical:

1Build a 25-second vertical (9:16) short from this 5-shot list.
2Seedance: generate each shot as a 5-second moving clip, back to back.
3Put the hook text over shot 1 for the first 2 seconds, large, high contrast.
4ElevenLabs: read the narration in one locked voice, word-by-word captions, bottom third.
5Score it with a low, tense ambient bed that never covers the voice.
6Hold the final shot on the open loop. Export for TikTok and Shorts.

It locks recurring characters and settings as reusable assets. Cast once, change the plot, ship forever.

Step 5. PUBLISH: Blotato Posts It Everywhere

One call to Blotato pushes the vertical to TikTok, YouTube Shorts, and Instagram Reels at once. The agent writes titles and tags from the format tag, so one metadata skeleton runs the whole series.

It fires the post while I sleep and logs the post ID for the last step.

Step 6. VIRAL VIEWS: The Loop That Makes It Smarter

This step separates an agent from a workflow. 48 hours after each post it runs:

1. Pull view counts on every clip it shipped, per platform
2. Compute the outlier score of its own videos against its own median
3. Flag anything that beat its median by 10x
4. Add the winning format to the top of tomorrow's idea pool
5. Downrank any format that died twice
6. Rewrite its own idea prompt with the week's proven shapes
7. Send me one Telegram line: which format won, what it is scaling next

The agent's own hits become its inputs. The formats that work for your account rise on their own, and the idea list gets better every week without you touching the prompt.

Setup: 5 hours once. Runs indefinitely after.

How This Makes Money

• Platform payouts: 4 to 8M monthly views clears $2,500 to $7,000/month.
• Channel sales: a proven format at 100k followers sells for $4,000 to $12,000.
• Done-for-you shorts: a 20-short retainer clears $3,000 to $6,000/month, cost pennies.
• Research service: the outlier tracker as a monthly report, 60 subs at $29 is $1,740/month.
• Sell the agent: the Kimi config plus prompts on Gumroad at $79, sold while it runs your channels.

The Numbers

Month 1: build, $0 Month 2: $700-1,400 Month 3: $2,400-4,000 Month 6: $6,000-9,000 Month 9: $9,000-13,000 Month 12: $13,000-18,000+

The loop compounds. Every week the agent's idea pool gets sharper and its own hit rate climbs.

The Cost

Kimi K3 $12/month: $30/month in API Seedance: $30/month ElevenLabs: $22/month Blotato: $29/month

Total: $111/month in. $14,700 out.

The Bottom Line

Most people run the tools by hand, ship one video, and hope.

One agent finds the winning format, builds it, posts it everywhere, reads the result, and points itself at the next hit before you wake up.

Four services. One agent. You approve. It ships.

comment "AGENT" and I'll send the Kimi config and the prompt stack.