How I Created My First AI Video

My first AI-assisted video went through several versions before I liked it.
The voice mispronounced technical terms. The first character did not look right. A more realistic character looked better as an image, but its animation still felt fake. The narration needed a faster pace, and the layout had elements I eventually removed.
The final result was a 48-second, character-free explainer about using Chrome DevTools MCP for SEO checks, complete with AI narration, animated graphics, background music, captions, and an Instagram cover.
This is how I created it, what I changed, and what I would repeat for my next video.
Watch the finished video on Instagram
What I wanted to create
I wanted a short vertical video that explained a useful SEO workflow in plain English.
The topic was Chrome DevTools MCP: a way for compatible AI assistants to interact with Chrome and help inspect websites. The project supports browser automation, console and network inspection, and performance analysis. Those capabilities made it a useful subject for an SEO explainer. Source: Chrome DevTools MCP documentation
The opening idea was simple:
What if your AI could open Chrome and spot SEO issues for you?
The video would introduce the tool, show a configuration example, explain the checks someone could request, and close with a save-and-follow prompt.
I initially aimed for 60 seconds. After reviewing the pacing, I chose a faster delivery and a 48-second final version.
The tools behind the video
I used an AI coding assistant to help build and revise a local video workflow. Different parts of the process used different tools.
| Tool or component | How it contributed |
|---|---|
| AI coding assistant | Helped refine the script, create the project files, build graphics, and implement revisions |
Kokoro through kokoro-onnx |
Generated the English narration locally |
Fenrir voice, am_fenrir |
Provided the male American English voice I selected |
| HTML, JavaScript, and Canvas | Defined the layouts, diagrams, captions, and animation |
| Puppeteer and a local Chromium browser | Rendered the programmed visuals into frames |
| FFmpeg | Encoded the frames and combined them with the audio |
| AI image generation | Created character experiments and the final thumbnail |
| A local music-generation script | Created the instrumental background track |
| Whisper | Helped with a pronunciation transcription check during revisions |
The voice setup used the kokoro-onnx project, which runs Kokoro text-to-speech with ONNX Runtime.
I also had tools such as yt-dlp and HyperFrames installed in the workspace. However, the final video’s rendering path was the custom Canvas, browser, and FFmpeg workflow. Installing a tool and using it to produce the final video are different things.
The setup reused existing Python and Node.js installations, while project packages, models, assets, and working files stayed inside my video workspace. Image generation was a separate AI service, so this was not an entirely offline workflow.
1. I established a reusable visual style
Before finishing the video, I wanted consistent fonts, colors, and layouts that I could reuse.
The style settled on:
- A midnight navy background:
#101827 - Cyan accents:
#65D8ED - Lime highlights:
#D5FA52 - Off-white text:
#F7F6F2 - Bold Montserrat typography
- Vertical 1080 × 1920 output
I also set some editorial preferences: English narration, no em dashes, clear pronunciation of technical terms, and script approval before final production.
Having a written style guide made revisions more consistent. When I changed the presentation direction later, the colors, typography, and overall identity still held together.
2. I approved the script before rendering
I supplied a script with six sections: hook, introduction, connection setup, SEO demonstration, advanced checks, and closing.
The first draft needed tightening for a short video. We revised it to roughly 125 words and kept the central idea intact.
One useful correction was adding that setup varies by AI tool. A generic configuration example can be helpful, but it should not imply that every application uses exactly the same setup process.
This was the JSON example shown in the video:
{
"mcpServers": {
"chrome-devtools": {
"command": "npx",
"args": [
"-y",
"chrome-devtools-mcp@latest"
]
}
}
}
The example matches the project’s documented MCP configuration. The official guide also lists Node.js, npm, and Chrome as requirements and directs users to client-specific setup instructions. Source: official setup guide
After reading the complete revised script, I approved it. The project recorded that approval against the script file so an accidental wording change would not silently become the final narration.
That approval step became part of the reusable workflow.
3. I chose a voice, then fixed its pronunciation
I reviewed male voice samples and selected Fenrir.
The voice suited the video, but technical abbreviations needed extra attention. In particular, AI and MCP did not consistently sound the way I wanted.
My pronunciation requirements were straightforward:
| Written term | Intended pronunciation |
|---|---|
| AI | Ay eye |
| SEO | Ess ee oh |
| MCP | Em see pee |
The solution was to separate the written text from the speech instructions. Captions could still display “MCP,” while the voice engine received an explicit pronunciation.
We also found a technical problem: a saved pronunciation mapping had been corrupted into question marks. Repairing the encoding was necessary before the speech engine could use it correctly.
This was a practical reminder to check both the audio and the configuration behind it. Repeatedly asking a voice model to “say it correctly” would not have fixed corrupted text in a file.
4. I increased the narration speed
The initial delivery felt slow to me, so I requested a voice speed setting of 1.25×.
That change also required adjusting the scene timing. Simply making the voice faster while leaving every scene untouched would have introduced longer gaps between sections.
The workflow measured the generated narration and redistributed the scene boundaries around it. The final video ran for 48 seconds, with the approved wording preserved.
For this project, the faster version felt better. I would still review the pace separately for a tutorial with more code or a concept that needs extra explanation.
5. I tried AI characters and eventually removed them
The character was the most obvious example of an idea that did not survive the editing process.
I first wanted a Punjabi male presenter. We tried an illustrated character, then moved toward a more polished image with a burgundy turban, charcoal blazer, and ivory shirt.
The second character looked better as a still image. But movement was another problem.
At first, the character had very little motion. Its hand also pointed in a direction that made sense beside some graphics but looked awkward when the MCP configuration appeared.
We added arm gestures, head movement, and a scene-specific direction change. Even after those revisions, the animation felt artificial to me.
So I asked to remove the character completely.
That decision improved the direction of the video. The screen could focus on the concept being explained, with more room for readable text and diagrams.
6. I rebuilt the scenes as motion graphics
Removing the character required more than deleting an image. The earlier layouts had reserved space for a presenter, so the remaining elements needed to be repositioned and enlarged.
The final version used:
- A browser-style opening panel
- A diagram connecting the AI tool, MCP server, and Chrome
- A readable JSON configuration panel with changing highlights
- Sequentially highlighted SEO checks
- Performance and network graphics
- A prioritized-report layout
- A clear follow-and-save closing card
I also removed the top branding pill, progress bar, and scene counter from the earlier design.
The audit section was labeled illustrative. It explained the workflow using graphics rather than presenting invented findings as a real audit of a particular website.
That distinction matters when creating technical content. An animation can show how a process works without pretending that it is a screen recording of an actual result.
7. I added music and captions
The video needed more energy, so we added an original instrumental background track generated by a local script.
The music was mixed at a lower level and automatically reduced under the narration. This helped keep the voice clear while giving the video some rhythm.
I supplied reference reels during the process, but the final background track was not copied from those videos.
Captions were also added in short, readable chunks. Their timing was approximate within each narrated section, rather than a precise word-by-word transcript alignment.
For future videos, more accurate caption timing is one area I can continue improving. For this version, the priority was keeping the words readable and the screen uncluttered.
8. I reviewed the output and kept the files organized
The workflow generated preview frames so the layouts could be checked before the final render. The finished file was also checked for duration, video and audio streams, and decoding errors.
The final output was:
| Property | Final video |
|---|---|
| Duration | 48 seconds |
| Resolution | 1080 × 1920 |
| Frame rate | 30 fps |
| Video format | H.264 in an MP4 file |
| Audio format | AAC |
| Presentation | Character-free explainer graphics |
| Voice | Male Fenrir, with a 1.25× speed setting |
Technical checks helped confirm that the file worked. They did not replace my own judgment about whether the voice, pacing, or visuals felt right.
I also settled on a folder structure for future projects:
My Video Editor/
├── raw/ Original footage
├── references/ Inspiration and reference videos
├── projects/ Scripts, audio, graphics, covers, and previews
└── export/ Completed final videos only
New exports received versioned names so earlier versions stayed available. That was useful because the changes were substantial, especially when I moved away from characters.
9. I created a matching Instagram thumbnail
Once the video was ready, I requested a cover in the same visual style.
The headline was:
Let AI Check Your SEO
The thumbnail used oversized white and lime text, a navy background, a cyan subtitle, and a browser-style graphic showing titles, canonicals, and performance.
It followed the final presentation direction: no character, one clear topic, and a simple visual hierarchy.
I liked the result and asked to keep that approach for future videos. The specific headline will change with the topic, but the visual identity can remain consistent.
What I learned from creating my first AI-assisted video
The most useful part of the process was being able to revise individual pieces. I could change the voice speed without rewriting the message, replace a visual style without abandoning the script, and remove a character while preserving the rest of the project.
These are the lessons I am carrying forward:
- Approve the message early. A clear script gives the rest of the production a stable foundation.
- Check technical pronunciation separately. Short acronyms can be more troublesome than long sentences.
- Judge characters in motion. A good generated portrait does not guarantee convincing animation.
- Use the available screen space well. Larger diagrams and code panels were more useful for this topic than a presenter.
- Adjust timing after generating speech. Word count alone does not tell you how a voice will deliver a sentence.
- Keep reusable guidelines. Saving preferences reduces repeated corrections across future projects.
I am not attaching a precise production cost or time-saving claim to this experiment because I did not track those consistently. It also took several rounds of feedback, so describing it as a one-click result would give the wrong impression.
What I did end up with was a finished video and a more repeatable workflow for the next one.
The workflow I will use for my next video
My next project will start with the decisions this one helped me make: a focused script, English Fenrir narration, clear acronym pronunciation, character-free graphics, restrained background music, and a matching thumbnail.
I will still review the script before production and judge the result after rendering. New topics may need different pacing or layouts, and I expect the style to keep evolving with feedback.
If you are creating your first video using AI, start with one useful idea and a short script. Review the voice early, check the visuals at phone size, and be willing to remove an element that is making the explanation harder to follow.
I will be sharing more of these practical experiments on DgTrendz, including the changes that worked and the ones I decided to leave behind.