How we got 47,200 views from an AI Avatar Reel edited with Claude

📢 One AI generated Reel on our brand account has 47,200 views. Nobody on our team is in it.

↳ No camera.

↳ No studio.

↳ No face on screen.

↳ No actor to book, brief or pay.

The only thing we recorded was a screen.

Here is what makes that number mean something. Look at the Reels sitting next to it on the same account, published to the same audience, in the same weeks.

— "Our AI avatar video in the Distribb.io instagram account"

Those numbers again: 607, 1,933, 2,188, 2,365, 2,969, 3,547, 6,379.

And then 47,200.

Check out the actual video here: instagram.com/reel/DZrwttgT2ZC

Fair credit before anything else. Florian, my co-founder, made that one with Arcads. But I built an open source skill that accomplished the same.

See the open source skill here.

I also rebuilt the workflow inside Distribb afterwards, so our Accelerator members get AI generated videos created and distributed for them as a done-for-you service, and everyone else can run the same steps themselves through Claude.

We tried to repeat it and got 607 views

This is why getting the avatar's expressions right is so important

We repeated the video with a different avatar...

It only got 607 views

Same script. 78 times less reach.

The facial expressions and the movement it produced were slightly off. Not broken. Not cartoonish. Slightly. The kind of wrong you cannot catch in a still frame and cannot miss in half a second of video.

None of this is new. Masahiro Mori predicted it in 1970, in the essay that gave us the phrase "uncanny valley."

Everyone knows the graph. Almost nobody quotes the part of the argument that matters here, which is entirely about movement:

> "Its presence changes the shape of the uncanny valley graph by amplifying the peaks and valleys."

His point is that a still figure sitting near the bottom of the valley reads as merely odd, while the same figure in motion reads as eerie. His example is a prosthetic hand: passable at rest, unsettling the moment it moves.

Then he makes it blunter and contrasts a corpse with a zombie. Same thing, one of them moving.

Our still looked fine. It moved, and it fell in.

Mori was writing about robots in 1970. He was describing AI videos 56 years early.

Two gates, and both have to open

↳ Gate one: the avatar has to move like a person for one second.

↳ Gate two: the script has to earn the other 33.

Miss either one and you get 607.

Instagram tells you why itself. Open any Reel and go into the video-specific analytics.

— "Screenshot of Instagram engagement where they specifically list what impacts your views the most"

"Rates are listed in order of importance to reach."

Skip rate is first. Above shares, above saves, above comments. Instagram is telling you in its own dashboard that the most important thing your Reel can do is simply not get skipped.

An avatar that moves wrong gets skipped in the first second. A Reel with a high skip rate never gets distributed. The script you spent all that time on never gets heard by anyone. That is how a video with a proven script does 607.

Now the other half, because the avatar cannot do this part for you:

Answer engines never see your avatar.

ChatGPT, Perplexity and Google AI Overviews read your transcript, your caption and your on-screen text. They do not watch the pixels. I looked into this carefully before deciding how much to spend on avatars in our own done-for-you pipeline, and it changed the answer.

So the avatar's entire job is the first second. Everything after that, the watch time, the comments, the ranking, the discovery a month later, is the script.

Our avatar Reel went up on June 17. When I recorded my video about it a month later, the most recent comment was 19 hours old. It was still being found. No avatar does that for you. The script does.

Instagram is a search engine now, and Meta lets Google and Bing index eligible public Reels from professional accounts belonging to adults, so one Reel gets found in the feed, in Instagram search, in Google, and inside ChatGPT when someone asks a question you happened to answer.

If you have not read the playbook behind the script, it is here: Claude Helped me get 588,000 Views With Instagram Reels AI Can Cite (9-Step Playbook)

The 47,200 view script, word for word

I pulled the exact audio out of that Reel. About 34 seconds, 119 words, and not one of them is filler.

> "Claude can now get your customers from Google on autopilot. Instead of paying an agency $5,000 a month, you install one skill that turns your agent into an SEO lead machine. It researches what your buyers actually search for. It studies your competitors and finds the content gaps they're missing. It writes SEO articles built on original research, then publishes them straight to your site. It gets you backlinks from real businesses. And here's the part an agency can't touch. It gets you showing up in AI search too. So when someone asks ChatGPT what to use, you're the answer. If you want your agent to build your entire distribution engine, comment SEO and I'll send you the Distribb Skill."

Read it again as five beats, because this is the actual template.

1. The trend plus the outcome. About 3 seconds.

> "Claude can now get your customers from Google on autopilot."

2. The pain and the mechanism. About 6 seconds.

> "Instead of paying an agency $5,000 a month, you install one skill."

A named price and a named alternative. The viewer knows exactly what they are being saved from.

3. The proof stack. About 15 seconds.

Five claims, one line each. Researches your buyers. Studies your competitors. Writes the articles. Publishes them. Gets the backlinks.

Every one of those lines is a screen change. That is the whole reason the middle of the video does not sag.

4. The thing a competitor cannot say. About 6 seconds.

> "And here's the part an agency can't touch. It gets you showing up in AI search too."

5. The keyword CTA. About 4 seconds.

> "Comment SEO and I'll send you the Distribb Skill."

One honest note: 34 seconds breaks my own 15 to 30 second rule. The proof stack earns the extra few seconds. It is the only rule I would let an avatar video bend, and I would not push it to 45.

And remember: the 607 video had this exact script.

The layout

Three stacked pieces, and the order matters:

↳ Screen recording on top. The proof.

↳ Caption in the middle. The claim being made right now.

↳ Avatar at the bottom. The person saying it.

Look at the comments in that screenshot. florian.darroman: SEO. borja.obeso: Seo. mr.amarisoul: SEO. hockey_ice_time: SEO.

That is the keyword CTA doing its job. Every one of those is a DM that gets sent automatically.

The workflow

1️⃣ Start with a real face

Use a photo of yourself, or of a real person who has given you consent.

Not a fully invented one.

These models were trained on real humans, so a real reference image animates more naturally than something generated from nothing. You are not just picking a look here. You are giving the animation model something realistic to move.

If you do not want your own face out there, ask someone. Get the consent in writing.

2️⃣ Generate the avatar with GPT-Image-2

Feed the reference photo in and generate your presenter.

You can ask it to change the face enough that it is no longer recognisably the person you started from, while still looking like a real human being.

Keep the framing boring. Head and shoulders, plain room, decent light. You are making a talking head, not a portrait.

Do not fall in love with the still. The still is not the product.

3️⃣ Write the script before you open any video tool

Use the five beats above, then check it:

↳ Does the first line name a trend and an outcome?

↳ Is there a number in the first 10 seconds?

↳ Does every single claim have a screen you can show?

↳ Is there one CTA with one keyword?

↳ Can you say it out loud in under 35 seconds without rushing?

Claude will write this for you. Give it the transcript of a Reel that already worked, tell it to keep the structure and change the angle.

A good script will not rescue an avatar that moves wrong. We proved that at a cost of 46,593 views. But a bad script wastes a good avatar, and good avatars are the expensive part.

4️⃣ Animate with Seedance 2.0, and this is where it is won

Feed the avatar image in, animate it, then watch it properly.

Not scrubbed. Not paused. At full speed, on a phone, the way it will actually be consumed. Watch the first second on its own about ten times.

AI has quirks. The mouth drifts. The eyes go dead between sentences. The head holds too still, then snaps. An eyebrow does something an eyebrow does not do. None of that shows up in a frame grab.

Regenerate until the movement stops bothering you. This is a retry loop, not a one-shot, and you should budget for it.

This is the step where the 78x lives. Everything else in this article is cheap by comparison.

One gotcha nobody warns you about: Seedance runs through third party providers like Replicate and fal, and they can get funny about photographs of real humans. If your generation keeps getting rejected, that is usually why, and you will have to work your way around it.

5️⃣ Record real screen footage

This is the half that makes it credible.

Florian recorded his own screen for that Reel. It is the reason it does not feel like an ad.

Record far more than you need, in long clean takes, so an agent can cut it later:

↳ The product page.

↳ The Skill's GitHub repository.

↳ A command running.

↳ The output appearing.

↳ An article being published.

You can get Claude to drive the browser and record this for you. It will not be as fluid as a human doing it by hand. Prompt it well and it gets close.

6️⃣ Add captions

Most Reels are watched on mute.

If your words are not on screen, the script you just spent all that time on does not exist.

7️⃣ Edit in the Edits app

Avatar bottom, captions middle, screen recording top, cut on every claim.

5 to 10 minutes.

I use Edits because it is Instagram's own app, and I have a suspicion the platform likes videos made in it. That is a theory. I have no proof.

8️⃣ End with a keyword CTA

"Comment SEO and I'll send you the link."

Distribb handles the reply for me. ManyChat does the same job.

The automated reply delivers the thing you just demonstrated, instead of making the viewer hunt through your bio for it.

9️⃣ Disclose that it is AI

Label it. Instagram gives you the toggle.

You are not going to out-clever a platform that can detect this at scale, and trading your reach for two seconds of honesty is a bad deal.

🔟 Turn the winners into a system

↳ Which angle produced the reach.

↳ Which hook kept showing up in the best ones.

↳ Which keyword brought people with buying intent.

↳ Which comment revealed the next video to make.

Then make another version of the winner rather than a brand new concept.

And when a generation moves well, keep it. Reuse it. That is the part we learned the expensive way. A presenter that clears gate one is an asset, not a setting you regenerate on a whim.

The whole workflow lives in a markdown playbook inside the Distribb Skill. Bring your Seedance and OpenAI keys and Claude runs it.

Where these videos fail

I am not going to pretend this always works.

↳ The animation is not a detail. Same script, different generation, 607 versus 47,200. The movement is the product.

↳ The still frame lies to you. You will approve it looking at a picture. Your audience judges it in motion, in one second, with their thumb already moving.

↳ Near-duplicates get punished. Generating 50 videos from one template is how you get demoted. Change the angle, not just the words.

↳ The avatar still cannot rank for you. It buys the first second. The transcript buys everything after.

What this actually replaces

Not your face. If you can record yourself, record yourself. It converts better, it is free, and a real human clears gate one automatically.

What it replaces is the excuse.

"I did not have time to film today."

"I do not like being on camera."

"My co-founder is in another country."

Those are the reasons most people publish nothing for three months.

Florian is in a different country to me. That Reel went out anyway, and 47,200 people saw it.

If you want to automate the research, the SEO keyword selection, the script and the message delivery behind this system, try Distribb here.

And if you want us to make these for your company, send me a partnership request.

Thanks for reading!

Borja

P.S. Curious about how Distribb becomes your AI SEO agent? Here's my Co-Founder Florian explaining it: