You can't escape it.
Literally EVERYWHERE you look on your feed is filled with something about "jEv is iNsAnE"... and it's generated tens of millions of views on X in less than a week.
Real.
I'm guilty of it too.
I've posted about it TWICE in the last few days.
Everyone else riding this wave, has shown a slop demo (that they actually created solely with Claude), then shared a take on something they don't \actually\ understand.
But not a SINGLE person, other than me, has ACTUALLY shown it functioning in their tool.
(something like this):
So what I need you to do is: stop believing what you see on your feed, and pay CLOSE attention to what I'm about to show you... because my team and I have actually built Jev into our SaaS, and are running experiments with it EVERY SINGLE DAY.
And if you're in content marketing, you probably dream of Jev:
- finding hooks, formats, and angles that are blowing up in your niche before your competitors do
- turning this data it into scripts/briefs, proven by performance data, that you can record to.
So that's what this article is gonna teach you.
What's inside
- What Jev actually is, in plain English
- How to set up a content research agent that scrolls TikTok for you
- What every video has to pass before it reaches you
- How Jev watches, analyzes, and judges viral videos
- The experiments we're running w/ Jev vs. our current judge
- Jev + Virlo = viral scripts & briefs
- FAQs
- How to go viral with this
What Jev actually is
Before any of this connects to content marketing, you need to actually understand the concept everyone's hyping up.
Because once you get it, the content marketing use case writes itself.
Every model you've used before this works the same way: Claude, GPT, Gemini.
You ask it something, it writes you an answer, word by word.
Even a yes-or-no question gets the full essay treatment:
- The AI model thinks out loud.
- Then it burns tokens and time on a decision that should take a fraction of a second.
Think about how your own brain actually works.
Two systems, running side by side.
One is slow and deliberate, working a real problem out on paper, whilst the other processes information instantaneously... like "reading the room", the second you step through the door.
Every AI model until now dreaded on the first one - and you know exactly how it feels to use them.
Well?
Jev is the FIRST real instance we're seeing of instantaneous processing in AI.
The efficiency Jev brings to Virlo's solutions
@TypeSafeAI built it, founded by Diogo Almeida - a co-creator of ChatGPT and RLHF.
https://x.com/CompleteSkeptic/status/2099925682726002904
It takes a messy situation and a set of possible answers, and returns a decision, not a paragraph, with a confidence score attached.
Three ways you can ask it something:
- Choice → pick one option out of a list you define. "Is this a tutorial, a review, or a storytime video?"
- Score → rate something against a scale you define. "How strong is this hook, weak to viral-caliber?"
- Noul → answer a yes-or-no question, returned as a probability, not a flat true or false. "Is this video actually about the niche it's hashtagged under?"
Ask all three about the same video, in one request, and Jev answers every question in parallel. One question or fifteen, the response time barely moves.
Every answer comes back with a confidence score attached.
The confidence score returned by Jev, in Virlo
- High confidence?
Act on it.
- Shaky?
Escalate it, to a stronger model or a person.
Sounding confident and being right aren't the same thing, and that's exactly where Claude's hallucinations come from.
Jev got trained on the opposite: honest confidence.
Say 90%, be right 90% of the time.
(Routing a "does this match" call through a
full-reasoning model
is a bit like calling a surgeon to check if a door's locked.
Teeechnically capable... just the wrong tool for the job.)
And they're up to 200x faster, 400x cheaper than a comparable model on classification work.
"Jev doesn't create content for you... It judges. what you should be creating."
Here's why this is genuinely GAME-CHANGING for marketing in 2026, and why you need to get ahead before you're playing catch-up with your competitors...
Because when it comes to content research...
YOU ARE the robot.
- Constantly doomscrolling TikTok and Instagram for HOURS
- Trying to manually spot which hooks, formats, and angles are actually working in your niche.
- AND trying to turn all this manual data into content ideas that \might** just break 100 views...
A human can only perform at their best for so long.
The first 5-10 minutes spent scrolling & saving content always feel amazing...
Then fatigue sets in.
A brainrot video slips into the feed, you get distracted, or you just get bored, and the accuracy that was there ten minutes ago quietly disappears.
You know that desired level of performance doesn't come back until the next day.
And the cycle repeats itself day-by-day.
But there's a tool BLOWING up right now, that completely breaks this cycle.
It's called Virlo, and it's not your average AI social-listening platform... it actually does the research FOR you.
So here's exactly how it works:
- Virlo creates "content research agents" that watch your niche or industry 24/7, pulling real videos from a database of over 12.8 million viral videos across TikTok, Instagram Reels, and YouTube Shorts
- It watches, analyzes, and judges every single video, now via Jev's judgment layer, which I'll be breaking down below
- Turns the outliers into winning scripts and briefs, backed by proven short-form performance data
And you need to MOVE FAST, because we just crossed 100,000 users on the platform, with more jumping in every single day.
A 1-minute demo video on using Virlo to find viral content ideas in 3 prompts.
Don't know where to start?
All good - the next section is a full step-by-step guide on setting the whole thing up.
How to set up a content research agent that scrolls TikTok for you
Now here's one simple choice for you to make:
- Prefer a real interface... somewhere to click into the dashboard, browse your agent's research, see the videos yourself?
→ Set up your first content research agent at virlo.ai
- Prefer running it through an AI agent like Claude Code instead + Virlo's MCP?
→ Grab an API key at dev.virlo.aiThen pass these prompts to your agent:
PROMPT 1: Wire Jev into your pipeline
1Goal:2- connect my TypeSafe Jev API key so every video from my Virlo marketing agent3 gets judged before it reaches me45Keys:6- Jev (TypeSafe) API key: [PASTE YOUR TYPESAFE KEY]7- Virlo MCP server: https://dev.virlo.ai/api/mcp/mcp8- Virlo API key starts with virlo_tkn_910Process:11- walk me through connecting both12- once connected, confirm by running one real video through Jev and showing me the judgment
PROMPT 2: Set up your marketing agent
1Goal:2- set up a Virlo marketing agent that tracks what's outperforming in my niche34Process:5- ask me for my niche, my brand name, my competitors' names, and which platforms6 (TikTok, Instagram, YouTube Shorts)7- draft the keyword mix, then confirm it with me before saving89Save the agent ID when it's done.
PROMPT 3: Get Jev's judgment on every video
1Goal:2- run every video my marketing agent [AGENT ID] surfaces through Jev34Process:5- for each video, get Jev's judgment directly, not a description:6 1. Noul: worth scripting, 0 to 17 2. Choice: the hook type, the format, and the angle8 3. Score: how it performed against that creator's own baseline9 4. Confidence attached to each call10- only hand me the ones that pass the Noul threshold
📌Pro tip if you're using MCP/API: Simply passing this entire article as a Markdown to Claude will get you the strongest results...
(But what I've noticed from Stripe data, is a lot of our more experienced marketing teams use both the API and the web app.)
What every video has to pass before it reaches you
Now, what you've probably been waiting to see this entire article... here's how Jev ACTUALLY judges videos that you content agent finds.
I set up a content research agent for a 9-figure Ecom brand called "Bloom Nutrition", and asked one question: Which hooks, formats, and angles are actually working in this space right now?
And in less than 20 seconds, the agent returned 384-plus videos, already broken down by hook, format, and angle.
Just to clarify: that number isn't because only that many videos exist about Bloom.
It's actually because a lot more than that got looked at, and judged in our 12-million video database, before any of it even reached us.
Every video pulled into a research run needs Clearance first.
Two specific checks, not a vague "trust me bro" call:
- Does this actually match the niche?Judged off the caption, the hashtags, and the transcript, not the hashtags alone.
- Are the hashtags lying?A video can hashtag-stuff its way into every trending niche search while the actual content is about something else entirely.
Anyone who's manually scrolled a niche and opened a "perfect match," only to find an unrelated video wearing the right tags, already knows how often this happens.
Both are structured, yes-or-no-with-a-confidence-score decisions, a fast repeatable call applied to every video every time, instead of a person eyeballing hashtags and hoping.
A single Clearance check runs on roughly 7,000 input tokens for a full batch of 25 videos.
A fraction of a cent.
Cheap enough to run on every video, every time.
Run that same decision through a full reasoning model instead, across hundreds of videos a day per niche, and you're either paying a real, compounding cost for something that should be near-instant...
...or skipping it and just trusting the hashtags.
Most setups quietly do the second one.
Every video you're looking at from here already earned its place, so the hour you'd have spent second-guessing hashtags is an hour you spend scripting instead.
How Jev watches, analyzes, and judges viral videos
Once a video clears, it lands inside a content research agent.
One that already benchmarks 80 performance signals on every single video, side by side, to expose which hooks, formats, and angles are actually winning in a niche.
Six real categories.
Jev's judgment plays into each one differently, category by category.
Category 1: Viral Hooks Intelligence 💬
A hook gets scored against every other hook in the niche, never judged alone.
Jev's role sits upstream of the whole thing:
- A hook-performance comparison only means something if every video in that comparison actually belongs to the niche
- Feed it a hashtag-stuffed video and you're benchmarking a hook against a sample that was never real to begin with
Category 2: Viral Format & Topic Intelligence 🧩
Same logic, sharper.
- A format pattern (review vs. tutorial vs. storytime) only holds up as a signal if the topic underneath it is genuinely on-topic
- This is the category Clearance protects most directly, since topic-mismatch is exactly the second check it runs on every video
Category 3: Visual Intelligence 📸
VirloACTUALLY watches the video - Jev just takes Virlo's results, then combines it with the caption, hashtags, and transcript to make a decision.
- This category sits downstream of Jev's judgment, not inside it
- Clearance decides whether a video's worth analyzing at all. What the camera's actually doing gets tagged once it's through.
Category 4: Production Intelligence 🎥
Same relationship as visual.
- Clearance decides entry
- How it was shot, what setting it's in, gets tagged after
Category 5: Speech Intelligence 🎙️
The one category Jev actually touches directly.
- The transcript is one of the three real inputs it judges against
- And NOT ONLY the transcript, but the soundwaves of each video as well
- A video whose spoken content doesn't match its hashtags is exactly the pattern Clearance is built to catch
Category 6: Product & Sales Intelligence 💸
CTA type, CTA frequency, brand-safety tier. All tagged per video.
- Get this category wrong on a niche full of mismatched videos and you'd conclude the wrong CTA converts, because half your sample was never actually competing in the same niche to begin with
- Clearance is what keeps that comparison honest before it ever reaches this tab
Across all six, Jev doesn't generate any of these signals itself.
That's Virlo's job.
It decides which videos earn the right to be counted in the benchmark at all. Get that wrong, and every category above inherits the mistake.
Six real answers, on which hooks, which formats, and which angles are actually converting in a niche, each one trustworthy because the sample behind it was judged before it was ever counted.
The experiments we're running w/ Jev vs. our current judge
I'm not gonna lie to you and say Jev's running the whole show.
Because it isn't.
Jev's in what we call Calibration inside Virlo right now: matched head-to-head against the judge it might replace, on real traffic, while we watch the results before handing it the keys.
Two real checks so far.
We ran it against 44 real, human-labeled videos across 6 research intents.
Jev got 33 right. Our existing judge got 30.
Small sample. Real head-to-head. Jev came out ahead.
Bigger one: a live comparison across nearly 300 real videos where Jev and the existing judge actually disagreed, the exact population worth checking by hand.
Early days. Only a handful graded so far.
But on the first few, Jev was right 3 times out of 4.
Neither's a finished study.
Both are real. Both point the same direction.
📌 A live, measured test, still short of a finished migration, and right now Jev's winning it.
Shoutout Austin - absolute POWERHOUSE.
Jev + Virlo = viral scripts & briefs
Everything above gets you a filtered, ranked, tagged set of videos. Want to close the loop all the way to an actual script? Send your content research agent this next:
1Turn the highest-ranked result into an actual script, not2just a brief.34Goal:5- stop handing over a list of videos in no particular order6- get a beat-by-beat script modeled on a real, currently-7 winning video in this niche, ready to hand to whoever's8 filming910Process:11- take the top-ranked, already-tagged results12- group by how each one opens: a question, a bold claim, or13 a before-and-after14- group by the real hook, format, and angle already attached15 to each result, not a guess16- pick the best-fitting outlier for the brand and write a17 beat-by-beat script modeled on its real opening, pacing,18 and structure, adapted, not copied word for word1920Tool:21- Virlo2223Knowledge:24- a script needs the real hook and structure behind it, not25 just a link to a video that performed well26- adapt the winning structure, never copy it verbatim
A script built on a hashtag-stuffed video is a script built on nothing.
That's why Clearance has to run first.
FAQs
Is this just riding the hype, or is it actually live?
Both.
The trend is real, I'm part of it, and the build underneath it is also real. Calibration is a live test running on real traffic, not a finished feature.
Does this replace the research I'm already doing by hand?
It's aimed at the hour you lose scrolling a niche, where half of what you save turns out to be off-topic once you actually watch it back.
Is Jev the thing tagging hooks and formats?
No, and it's the most common mistake in every thread this week.
Jev decides what's worth tagging. Virlo's 80-signal panel does the actual tagging. Two different jobs - DO NOT LISTEN TO LARPERS
What happens if a video skips Clearance?
Back to the old problem.
Some genuinely relevant results, plenty wearing the right hashtags without being about your niche at all, and every one of the six categories above inheriting the mismatch.
Is this just for agencies running dozens of clients?
It's for everyone!
We have multi-client agencies, in-house marketing teams, and solo operators/marketers all building viral content engines on our platform.
Why publish this while Calibration is still running?
A live, real head-to-head where the new thing is winning is worth more than a finished announcement later.
Who knows just how amazing Jev + Virlo will be in the future?
What building this actually taught us
The biggest jump in accuracy didn't come from a bigger model... or a longer prompt.
It came from one sentence.
Early on, Jev kept treating every detail in a research brief as a hard requirement.
Format. Tone. Even audience description.
Most of that's a preference, not a rule.
A stricter version of the same question, one that mirrored the brief's exact wording back at it, only got 16 of 44 right on our test set.
One line telling it the topic was the actual requirement, everything else a preference, moved that to 33 of 44.
Same test set.
One sentence.
Roughly doubled the accuracy.
How to go viral with this
Everything above is stealable, right now, whether you're a solo creator or running ten clients.
- Stop scrolling by hand. Set up one content research agent per niche and let it watch TikTok, Instagram Reels, and YouTube Shorts on a schedule.
- Stop wasting review time on off-topic videos. Every result already passed Clearance, so what you're looking at genuinely matches the niche.
- Stop handing over a pile of links. Turn the top result straight into a beat-by-beat script or brief, using the prompt above, hook and structure already attached.
Do those three, and the fast judge everyone's posting about stops being a headline and starts being the reason your niche's next viral video came from your research, not your competitor's.
