We recorded a 28-second video on a phone, in a room with background noise, and turned it into a reel with captions, graphics, music, and our brand’s closing screen without opening a video editor. We promised it in the reel itself: if you commented BLOG, we would tell you how we did it. Here it is, with what went well and with what the machine could not decide on its own.
The two videos in this story
In the reel we talk about two pieces.
The first is the ad for Rey Cagón, an entertainment app we built: 43 seconds of animation with the app’s screens, a podium, notifications, and the call to download it. That video was not edited in any program. It was written as code with Remotion, a tool that draws every frame of a video from code, and it was put together by Claude Code, an AI agent that writes and runs programs.
The second is the reel where we explain the first one. That one does start from a recording: a person talking to the camera, a single take, with hesitations and pauses. This article is mostly about how that second piece was edited, because it is the closest to what any business recording with a phone does.
The method: a skill, not a magic button
For the edit we used a Claude Code skill, that is, an instruction manual the agent follows step by step. It is called edicion-de-video and its foundation was created by Erik Taveras, who shared it with his community (comunidad.eriktaveras.com). We adapted it to our brand and added what this video forced us to solve.
The idea behind the method is simple: the AI does not invent the video, it measures it. First it analyzes the material (where the face is, where the pauses are, what is said in each second) and only then decides where everything goes.
Step by step
1. Clean the audio before cutting
The recording had a constant low hum, from a fan or an air conditioner. Measured, the voice was only 5 dB above the noise: in the pauses there was almost no difference from the words. Because of that, the tool that looks for silences believed there were none.
We solved it with a noise reduction filter trained for speech (RNNoise). The separation went from 5 to 25 dB, and then the pauses showed up. The lesson is now written into the method: clean first, then look for the silences.
2. Remove pauses and hesitations
With clean audio, 9 stretches of pauses, false starts, and drawn-out “and…” were detected and removed. The take went from 28 to 21.5 seconds. It was also cut right after the last word: what came next was the person standing still looking at the camera, and in a reel that feels like a mistake.
All the cuts were made in a single pass, so the video was not recompressed once for every cut.
3. Transcribe word by word
The transcription was done by Whisper, a speech recognition model that runs on the same computer, without sending the audio to any service. The important part is not the text: it is the exact start and end time of each word, which is what lets the caption highlight a word right when it is spoken.
4. Fix what the AI misheard
This is where the person comes in. Whisper heard “Clark” where “Claude” was said, “la guía” (the guide) where “la IA” (the AI) was said, and “certificamos” (we certified) where “simplificamos” (we simplified) was said. None of those words sounds odd on its own, and that is why they are dangerous: a caption burned into the video gets published as if the author had said it.
The rule we follow: whatever the machine corrects by guessing gets checked with the author before publishing. In this case, all three were confirmed.
5. Place every graphic where it does not get in the way
Before putting a single piece of text on screen, frames were pulled from the whole take to know where the chin was, where the hand was pointing, and which areas of the screen Instagram and TikTok cover with their buttons. With that:
- The Rey Cagón ad starts full screen and shrinks to the corner right where the hand points when saying “this video”.
- The captions sit below the chin, never over the face.
- When the skill is named, a card shows its name and its author, which is what the voice does not say.
- When viewers are asked to comment BLOG, the word stays big and still long enough to type it.
6. Apply the brand
The first cut came out with generic colors and type. Our video handbook already had the rules: a single accent color, the brand typeface, text that comes in whole and not word by word, no abrupt zooms, and the official closing screen asking for a DM. They were applied in the second version and saved into the method, so the next video comes out on brand from the first try.
7. Mix the audio by measuring, not by ear
The music from the Rey Cagón ad plays loud during the hook, drops clearly below the voice when the person starts talking, and comes back up at the close. The levels were calculated by measuring the actual loudness of the voice and the music, not by turning a knob until it “sounded right”. The final file comes out in the format Instagram and TikTok ask for in reels.
What the machine did and what it didn’t
| Task | The AI | The person |
|---|---|---|
| Clean noise, cut pauses, sync captions | Yes | Reviews the result |
| Transcribe | Yes | Fixes names and doubtful words |
| Place graphics without covering the face | Yes, by measuring the frames | Approves |
| What to say and how to record it | No | Yes |
| Which graphic adds something and which is noise | Suggests | Decides |
| Whether the video represents the brand | No | Yes |
The takeaway: the AI took the mechanical, repetitive work, which is exactly what eats the most time in editing. The judgment stayed human, and two rounds of comments from the author improved the video more than any technical adjustment.
What this means for your business
If your team records with a phone for social media, the hard part is not recording: it is editing every piece with captions, cuts, and a recognizable style. A method like this turns that editing into a repeatable process, with your brand applied the same way every time.
It works well when:
- You record to camera often and always in a similar format.
- You already have, or want to have, a defined visual identity.
- Someone reviews before publishing.
It is not worth it if you publish one video a month or if every piece is a different creative project. There, a person with an editor is still more practical.
And something that applies to any AI project: the result depends on the rules you give it. Ours came from getting things wrong first, and now they are written down. If you want to understand where an automated system ends and one that really uses AI begins, we explain it in automated bot or real AI.