Danny MorrisThe problem
I wanted to know whether one person could run a content pipeline without being on camera to publish, with fast turnaround and no drop in quality, using voice cloning, an AI avatar and a scripted workflow. Two content types ran through the same system: quick AI news shorts for reach, and AI tool tutorials for practical use.
Before building anything, I studied three creators running this format successfully in early 2025, for hook structure and pacing, thumbnails and visual identity, and voice and avatar quality.
What I built
ChatGPT Projects held the system: seven modules covering script structure, voice and tone, a tool tracker, retention rules, performance feedback, audience language and a library of script templates. ElevenLabs handled the voice clone, trained on nearly three hours of my edited voice. HeyGen produced the avatar. A thumbnail brief was written alongside every script, and an external editor handled the cut.
Every stage had a check before the video moved on: topic relevance, brand fit, script timing (hook in the first three seconds), retention rules, voice quality, lip-sync and whether the thumbnail matched what the video delivered. Anything that failed got rewritten or cut. Every edit came back to me for a final review before it went live.
Why I didn't fully automate it
There were tutorials at the time showing n8n workflows that took a topic in and pushed a finished video out, with no person in the loop. I looked at that approach and chose not to build it, because of quality control.
With the 2025 tools, ElevenLabs mispronounced words, HeyGen lip-sync drifted, and thumbnails sometimes promised things the script didn't deliver. Each needed a person to catch it. Without that, the realism broke and viewers focused on the AI instead of the content. The edit style, with quick cuts, moving text and a change every two to three seconds, was a deliberate choice to stop the avatar sitting on screen long enough to look artificial.
I would only take the checks out once the tools had proved they could be trusted.
What the data showed
I posted the same videos to TikTok, YouTube Shorts and Instagram Reels to see where the format would land. TikTok was where the signal was: 44,837 views across 28 videos, a 5.28% engagement rate and two breakouts. Run DeepSeek R1 Locally hit 10,400 views with 216 saves, and Free AI Text-to-Video hit 9,100.
YouTube Shorts was steadier but flatter: 30,464 views, about 1,088 per video, and no breakouts. Instagram ran the same content, but that period's data is no longer available from Meta.
Why I stopped, and what I learned
The tools and the editor were a fixed monthly cost, and views weren't turning into followers. The data wasn't improving on that over the test, so I stopped at 28 videos and kept the system and the skills.
Views without follows. People were saving the videos as useful, but had no reason to follow. The fix is to take a position on each news story, and I do that in every short-form brief now.
AI audio at its limit. Even professionally trained, the voice clone was at the ceiling of what the tech could do in 2025. Audio quality is the first thing I check now.
Uneven platforms. The same content did very differently on each platform. That told me about platform fit, not production quality.
What it led to
The videos from this project came up in the interview that led to my AI video work at Hance AI. That work included Lenovo, Intel and Ferrari F1 projects, and some of it was shown on the TechCrunch stage in San Francisco in 2025. Hance has since closed, but the work is still public.