this AI agent knows which video ideas actually get views

Your content calendar is blank again.

So you do what everyone does. You open TikTok. You find a competitor whose videos are working. You watch ten of them. You screenshot the three that hit hardest. You make your own version of the best one.

It gets 400 views.

You copied the winner and the copy died. Then you do it again next week, because the calendar is still hungry and copying competitors is the only research process anyone ever taught you.

Here's the brutal truth. You copied the topic. The topic was never the thing that worked.

The thing that worked was a mechanism buried inside the video. A curiosity gap. A counterintuitive claim. A before/after with no narration. That mechanism is transplantable. The topic is not. Copy the topic and you get a worse video about the same thing. Copy the mechanism and you get a fresh video that wins.

So I built an AI system that pulls competitor videos apart, finds the mechanism inside each one, and remixes those mechanisms into ideas that fit your brand. Then it ranks them. Then it writes the scripts.

It's called Short-Form Kit. Six skills. Open source. MIT. Free.

Competitor outliers in. Ranked, on-brand scripts out.

Get it free, open-source, below (scroll to the bottom).

Copying competitor videos feels like research. It isn't. It's the most expensive mistake in short-form.

Here's what actually happens when you copy a winner.

You see a video with 2 million views. You assume the 2 million came from the topic. So you make the same video about the same topic. But you're the second person to make it, the algorithm already spent its appetite for that topic on the original, and you brought none of the things that actually carried it: the hook shape, the pacing, the specific tension it resolved.

You shipped a cover band. Cover bands don't break out.

There's a worse version of this trap, and it's the one that quietly drains budgets. You copy a video that was never an outlier in the first place. It had 2 million views because the creator has 4 million followers and every video they post clears a million. That video underperformed. You just copied a loss and called it research.

That's the move nobody questions:

This isn't a creativity problem. It’s not even a production problem.

Nobody decoded the video before they copied it.

Nobody asked which part of it was repeatable. Nobody checked whether it beat the creator's own baseline or just rode their follower count. Nobody pulled the mechanism out and separated it from the topic. So the calendar got filled with covers of other people's covers, and the feed does not reward that.

The fix is a process. Decode, don't copy.

Short-Form Kit is a six-skill Claude Code plugin that turns competitor short-form videos into ranked, on-brand video ideas and finished scripts.

Three skills you interact with directly, three that run under the hood:

→ The idea engine. The orchestrator. Discovers outlier videos, runs the decode on every one, remixes the parts into ideas, ranks them, and hands the top picks to the scriptwriter.
→ The video decoder. Takes one video and breaks it into its reusable components: the root, the visual hook, the written hook, the spoken hook, and the three extractables. Runs standalone or inside the engine.
→ The script writer. Takes one approved idea plus your brand profile and writes a production-ready script: three-hook block (visual, written, spoken), beat-by-beat body with retention marks, CTA, and a shotlist.

It composes a few tools underneath. Virlo discovers outliers by niche keyword. ScrapeCreators pulls transcripts and cover frames. A Gemini pass adds visual analysis on the strongest outliers. You don't wire any of that by hand. The engine does.

It runs two ways. Keyword mode discovers outlier videos for a niche across Instagram, TikTok, and YouTube. Handles mode analyzes a specific list of competitor handles. Either way, the output is the same: ranked ideas, then scripts.

Works with Claude Code, Codex, OpenClaw, or any agent framework that reads skills.

This is the framework. It's the thing you steal from this email even if you never download the kit.

You never copy a video. You decode it first. Decoding means classifying a winning video into parts that are actually reusable, then rebuilding from the parts.

Step one: root it.

Before you analyze a video, classify it. A cooking tutorial and a cooking opinion piece share a topic and win through completely different mechanisms. If you skip this step you'll misread why the video worked.

The kit roots every video on a three-level ladder:

→ Level 1. Fiction, Non-fiction, or Lifestyle.
→ Level 2. Domain plus treatment. Tech + commentary, Food + demo, Money + opinion.
→ Level 3. The treatment sub-type: educational, argumentative, opinion-based, narrative, or demo.

An AI news short about a model launch roots to Non-fiction > Tech + commentary > opinion-based. A "build this in 60 seconds" video roots to Non-fiction > Tech + demo > demo. Same niche, different roots, different win conditions. Root first or you remix blind.

Step two: pull the three parts.

Every decoded video yields exactly three reusable components. Keeping them separate is the whole game.

The Viral Vector is what most people miss. It's not the topic. A curiosity gap in the first two seconds is a Viral Vector. "Counterintuitive claim stated as fact, proven inside the video" is a Viral Vector. A visual before/after that needs no narration is a Viral Vector. Those transplant onto any topic. The topic itself does not.

Copy the topic, ship a cover band. Borrow the Viral Vector, ship something new that runs on a proven engine.

Here's where the fresh ideas actually come from.

Most operators only watch their direct competitors. That's exactly why their content looks like everyone else's. Direct competitors give you saturated ideas. Everyone in your niche is already watching the same five accounts and copying the same five videos.

The kit maps competitors into four rings:

The extraction rule is the part that breaks the copy trap:

→ Viral Vectors are valid from any ring. A great mechanism is a great mechanism, no matter where you found it.
→ Interest Topics carry the most signal from indirect competitors. They share your audience, so the topics winning for them tell you what your audience is hungry for right now.
→ Formats carry the most signal from adjacent competitors. A format winning in a different niche proves the container works, divorced from any one topic. Bring it into your niche and you're early.
→ Direct competitors mostly give you format parity you already have. Low signal. Watch them for Vectors only.

Familiar but new. You borrow a format from outside your niche, a topic from inside it, and a mechanism from anywhere. The result feels native to your audience and looks nothing like what your direct competitors are shipping.

That's the difference between research and copying.

One more piece, because the kit refuses to get this wrong and most operators get it wrong every time.

An outlier is a video that beats its own channel's baseline. Not a video with big absolute views.

A channel that normally does 50K views posts one that does 500K. That 500K video is a 10x outlier. Real signal. Something specific carried it.

A channel that normally does 4M views posts one that does 2M. Two million views looks huge. It's a loss. The creator's own audience showed up at half the usual rate. There's nothing to learn from it except what not to do.

The kit measures every video against the creator's own recent median. In handles mode it pulls roughly 20 recent posts per creator and keeps only the ones beating that creator's median by 2x or more. In keyword mode it discovers videos already flagged as outliers relative to their channel. Every decode carries an outlier magnitude, 8x baseline or 3x baseline, and that ratio feeds the ranking.

Absolute views lie. Outlier magnitude tells the truth. Decode the 10x video, ignore the 2M-view loss.

The engine runs a ten-step pipeline. You talk to it in plain English. It handles the steps.

A few things worth slowing down on.

The decode is structured. Every video becomes a Decode Record with 16 fields: root, visual hook, written hook, spoken hook, the Viral Vectors, the Interest Topic and its type, the Format, the competitor ring, the outlier magnitude. Nothing is a guess. If the kit can't decode a video, it skips it and logs why instead of faking a record.

The deep pass is honest about what it knows. The top 5 to 8 outliers by magnitude get a Gemini visual analysis: pacing, cut rhythm, visual style, retention devices. Those decodes get tagged deep. The rest stay shallow, and any idea built on a shallow decode is flagged in the ranked table. The kit never pretends it analyzed footage it didn't.

Ideation borrows, it doesn't invent. Stage 6 pools every Viral Vector, Interest Topic, and Format from every decode and remixes them into 10 to 15 ideas. Each idea names exactly which Vector, Topic, and Format it borrowed, and states the brand angle. If the decode pool is too thin to support 10 real ideas, the kit makes fewer and says so. It will not fabricate a mechanism that wasn't decoded.

This is the part that separates a kit from a content firehose.

Most AI tools hand you 30 ideas. Thirty ideas isn't a plan. It's a second blank calendar.

Short-Form Kit scores every idea on two axes before it writes a single script.

Viral potential (1 to 10). The average of five sub-factors: Viral Vector strength, outlier magnitude, format repeatability, topic freshness, and hook strength.

Brand fit (1 to 10). The average of four sub-factors: voice match, audience match, content-pillar match, and authentic credibility.

Then it combines them:

Brand fit is a multiplier, not a tiebreaker. A viral-10 idea that scores 4 on brand fit lands at 4.0. A viral-7 idea that scores 9 on brand fit lands at 6.3. The on-brand idea wins. That math is deliberate. A video that goes viral for the wrong audience isn't a win, it's a distraction, and the rubric refuses to rank it like a win.

There's also a hard gate. Before any scoring, every idea is checked against your brand's anti-topics, the subjects you have decided you will never make content about. A match is an instant disqualification. A 10-viral idea on an anti-topic is still dropped, listed separately, and never scored. The kit will not tempt you with off-limits content because the numbers looked good.

Then the real gate. The kit presents the ranked table and stops. It does not script anything until you pick. Default suggestion is the top 3 to 5. You choose. You stay the editor. The agent is the research and drafting engine, not the decision-maker.

Once you pick, the script writer takes each idea and writes it: a three-hook block (visual hook, written hook, and spoken hook) where each complements instead of repeats the others, a beat-by-beat body, retention marks at every drop-off point (open loop, pattern break, callback, escalation, visual reset), a CTA in your voice, and a shotlist. Ready to shoot.

The rules baked into every stage.

→ Never call a big number an outlier. Outlier magnitude is measured against the creator's own baseline. A high-view loss is a loss.
→ Never invent an extractable. Ideas only remix Viral Vectors, Topics, and Formats that were actually decoded from real videos. Thin pool, fewer ideas, stated plainly.
→ Never infer footage it didn't watch. Shallow decodes are labeled shallow. Visual claims only appear when the Gemini pass actually ran.
→ Never script past the gate. You pick the ideas. The kit waits.
→ Never lose the receipts. Every view count and outlier ratio in any artifact traces back to a raw API response saved on disk. The run folder is auditable end to end.

If you have ever had a content strategy doc fall apart because someone caught one made-up number, these are the rules that stop that.

Here's the math nobody runs.

A video built on a copied idea and a video built on a decoded idea cost you exactly the same to make. Same shoot. Same edit. Same hour out of your week. The production bill does not care whether the idea behind the video was any good.

So the idea is the only variable that actually moves your numbers.

You're already paying to make the videos. The expensive part was never the kit or the research. The expensive part is shooting a week of content on ideas that were never going to land.

Short-Form Kit is free. The research it runs costs pennies per video. The only thing it changes is whether the idea behind each video was decoded or guessed. Across a month of posting, across a roster of brands, that one change is the whole growth curve.

Run this kit if:

→ You ship short-form video and the calendar is always hungry
→ You're an agency or operator running video for more than one brand
→ You keep copying competitor videos and keep wondering why the copies die
→ You have a brand you can articulate (voice, audience, content pillars, anti-topics) and want ideas that actually fit it
→ You're already running Slideshow Kit or Creator Breakout Kit and want the video-ideation engine that feeds them

Not for you if:

→ You're a solo creator with one channel and a deep instinct for your own niche. Keep trusting it, you're winning
→ You already have a senior short-form strategist running post-level outlier analysis and rejecting off-brand ideas in writing. You don't need this
→ You want a tool that hands you 30 ideas and calls it a strategy. That's the second blank calendar, and it's not what this is

Full system free on GitHub:

You get:

→ The idea engine that runs the full ten-step pipeline from competitor discovery to finished scripts
→ The video decoder that breaks any short-form video into a 16-field Decode Record (root, visual hook, written hook, spoken hook, the three extractables, outlier magnitude)
→ The script writer that turns an approved idea into a production-ready script (three-hook block with visual, written, and spoken hooks; beat-by-beat body; retention marks; CTA; shotlist)
→ The decode-don't-copy framework: rooting ladder, Viral Vector / Interest Topic / Format split, and the four competitor rings
→ Two discovery modes: keyword mode across Instagram, TikTok, and YouTube, and handles mode for a specific competitor list
→ Outlier logic that measures every video against its own channel baseline, not absolute views
→ A ranking rubric that scores viral potential against brand fit, with an anti-topic hard gate
→ An approval gate so you pick which ideas get scripted before any script is written
→ An honest deep pass that labels shallow vs deep decodes and never claims visual analysis it didn't run
→ A traceable run folder where every view count and outlier ratio traces back to a saved raw API response
→ Graceful degradation so a missing key disables one mode instead of breaking the whole engine

Runs on Claude Code, OpenClaw, or any agent framework that reads skills. MIT license. Fork it.

No upsell. No catch.

Stop copying competitor videos. Start decoding them.

Go big,

Matt

P.S. Short-Form Kit is the video-ideation engine. If you're already running Slideshow Kit, Creator Breakout Kit, Brand Shoot Kit, or Outcome Kit, this is the kit that decides what short-form video to make in the first place. Creator Breakout sets the creator strategy. Short-Form Kit decodes the outliers and writes the scripts. Slideshow Kit ships the carousels. Brand Shoot makes the visuals. Outcome Kit tells you which ones worked.

P.P.S. Star the Short-Form Kit repo if this helps. It tells me to keep building.