How I made my first YouTube video with a team of Grok Bots

By Viktor Szabó · Published 23 October 2026 · Updated 23 October 2026 · Checked against tool docs on 22 October 2026

Here’s the video: Email Marketing Audit for Brevo with Claude (Free Prompt, Read-Only).

I want to show exactly how it was made, including what went wrong. If you’re thinking about making videos with AI agents, this is what it actually looked like for me the first time.

The team: who did what

I run a small team of Grok Bots. Each one has a job, and none of them publishes anything without my yes. These are the ones that worked on this video:

My part was approving, or asking for a change. I didn’t edit anything myself. Every approval is written down as one line in a shared log, with the time and where the file lives. That log is where every time in this post comes from.

The steps, in order

1. The script (approved 00:22)

My agents turned a blog post draft about the same topic into a script: 1,411 spoken words across nine chapters, with the visual notes kept separate from what gets read out. I approved it at 00:22 with one small fix.

Before the voice was final, a claim check (a fact check my bots run on anything that describes a product) caught a line that made my team’s work sound like something I had done myself. It said the example was a report “that I pasted in” and “I’ve shortened it”. The fix now says it’s a real test run “made with this setup”. I approved that change at 01:00.

2. The edit plan (approved 00:36 and 00:37)

Video Editor wrote a long edit plan: 51 shots, the graphics, the music, the captions, and strict privacy rules for the screen recordings. I approved four big decisions:

• Nothing gets connected. No Brevo connector is added to my Claude account. The video shows Claude’s real directory page up to, but not including, the Connect button, and shows the login steps as text cards.
• A short real clip in Claude. Claude runs the audit prompt on a made-up report pasted into an incognito chat, with “Fictional list, not real numbers” on screen the whole time.
• An animated avatar. My illustrated avatar, instead of my face.
• YouTube’s altered or synthetic content box ticked, because the voice and the avatar are AI-made.

A minute later I picked the music mood: warm lo-fi, 80 to 90 BPM, instrumental. Video Editor found a CC0 (public domain) track, “Glad To Be Stuck Inside” by HoliznaCC0, and saved the licence page with the project. Its first choice of music site blocked access from its computer, so it switched.

3. The narration (approved 01:00 and 01:07)

Short Video Creator generated the narration with my ElevenLabs voice clone, one file per chapter. The final narration runs 8:56.

Two things needed fixing:

• One line kept coming out as “strongest” instead of “stronger”, in two full takes and a patch. The bot spliced in a short correct take.
• I asked for the spoken Gumroad link to come out of the ending. It now ends “The link is in the description.”

The bot checked each file with a transcript, word by word, against the script. I listened to the final version and approved it at 01:07.

4. The animated avatar (approved 01:55 to 02:27, delivered 03:25)

There was no talking version of my avatar, only still poses. Video Editor tested options and recommended Hedra Character-3 for lip sync and body animation, at about $34 for the full 8:56 at 1080p.

Then came a small surprise: I approved adding $35 to Hedra, but Hedra only sells $20, $50 or $100 API packs. So I approved the $50 pack instead. It was paid on my card at checkout (HUF 16,902.60, no tax), with auto top-up off.

I approved a 10-second test look, then the full render in 3:4 at 1080p. It came back at 03:25: $33.51 charged, $15.86 left in the wallet.

5. Five cuts (final cut approved 08:12)

Video Editor built the video in Remotion (video made from code) and checked each version before showing it to me:

• v1: the first full assembly.
• v2: the real screen takes, zooms and highlight rings. Some screen shots looked frozen, so they got a slow drift.
• v3: smoother push-ins on the still screen shots, and clearer wording on the card next to the Claude clip.
• v4: my ask. The avatar gestured away from the text, so I asked for it to be mirrored in the seven full-size avatar shots. Now it gestures at the words.
• v5: my ask. I wanted the Gumroad link off the screen at the end. It lives only in the YouTube description now.

The checks were thorough: loudness measured on the final file, lip sync checked at several points, every frame of the screen clips searched for my name, and the “Fictional list” label confirmed on every frame where example data shows. I approved v5 at 08:12.

6. YouTube (08:16 to 09:50)

• The channel. I wanted the video on a new channel tied to my new Google Workspace email, but that account wasn’t eligible for YouTube yet. So at 08:21 I chose a new brand channel, Viktor Szabó (@itsviktorszabo), to move over later.
• Upload. At 08:27 Video Editor uploaded v5 as Private, with the altered content setting on.
• Words. SEO & AEO Desk wrote the title, description, chapters and tags. I approved them at 09:36, with the main title “Email Marketing Audit for Brevo with Claude (Free Prompt, Read-Only)” and a clean Gumroad link.
• Links. Clickable links in the description need an ID or video check on YouTube. I chose to skip it for now, so the link is plain text.
• Thumbnail. Video Editor made three options. I picked option 2, “Find what’s broken”, at 09:48.
• Publish. At 09:50 I said “Publish it”, and it went from Private to Public.

Tools and real costs

So the new spend for this video was one $50 Hedra pack, and the render used $33.51 of it.

From my first approval (the script, 00:22) to public (09:50) took about nine and a half hours, overnight. That’s the clock time between two log lines, not a measure of how long the work took or what it would take a person.

What went wrong, and how it was fixed

• A line that oversold my part. The claim check caught it before the voice was final.
• A word the voice kept getting wrong. “Stronger” came out as “strongest” three times. Fixed with a short spliced take.
• A spoken URL. Long links are hard to hear. I had it cut and pointed to the description instead.
• Hedra’s pack sizes. I approved $35, but the smallest pack that covered the render was $50.
• Screen shots that looked frozen. Fixed with slow push-ins in v2 and v3.
• An avatar facing the wrong way. Mirrored in v4.
• A link on screen. Removed in v5.
• A YouTube account that wasn’t eligible yet. I chose a brand channel instead.
• Plain-text links. I skipped YouTube’s verification for clickable links for now.
• A music site that blocked access. Switched to a CC0 track.

None of these were big. But each one needed a decision from me, and I’d plan for all of them next time.

A checklist you can copy

Making a YouTube video with AI agents: my checklist

1. Write the script first and get it fact-checked, especially any line that says who did what.
2. Keep spoken lines and visual notes separate in the script.
3. Decide what must never appear on screen (accounts, emails, chat history) before anyone records anything.
4. Use fictional data for examples, and label it on screen the whole time it shows.
5. Listen to the full narration once before it's final.
6. Don't read long URLs out loud. Put them in the description.
7. Test the avatar or any paid render on 10 seconds before paying for the full length.
8. Check what the tool actually sells before you approve a budget.
9. Watch every cut. Ask for one change at a time.
10. Tick YouTube's altered or synthetic content box if the voice or face is AI-made, and say it in the description.
11. Upload as Private first. Add the title, description, chapters and thumbnail. Then publish.
12. Write down every approval with the time and the file. It makes a post like this one easy to check.