Codex + Hyperframes: Deconstructing Douyin Virals to Master the Self-Media Game

Previously, I tried using Codex and Hyperframes to deconstruct a viral video from a book-related Douyin account, and the sound effects received consistent praise from everyone.

Today, I am writing a long-form article to teach you step-by-step how to deconstruct viral Douyin videos and fix a workflow, so that you can also possess the ability to create videos using AI.

https://x.com/369Serena/status/2073012234184282287

1. Preparation

First, everyone must get their tools ready:

1. You need an Agent like Codex.

1. Then go to the plugin store to install two tools:

(a) Hyperframes: It creates videos in HTML format, which is very suitable for corporate presentations, data animations, or showcase videos.

(b) Remotion: Written in React, it achieves video-as-code. Once you fix a theme style, it is highly suitable for high-volume replication through coding. I see many bloggers using Remotion to create animations in the style of "Xiao Lin Shuo," which works very well with talking-head videos.

Additionally, you can install some Skills to assist in deconstruction. If you want to understand the structure of a viral video—such as how the first few seconds hook the audience, what the sound effects are like, and the rhythm between narration and visuals—you can deconstruct and replicate them through the following Skills:

1. Claude Video: Used to deconstruct the video process for easier replication in the next step.

1. Video Use: I highly recommend this for editing beginners. It acts like a "video editing director" that can call the appropriate tools based on your needs. For novices who don't know when to use which tool, it plays the role of a master director.

That concludes the preparation work.

2. Video Production Workflow

Creating videos with Codex + Hyperframes can follow this sequence:

1️⃣ Define video specifications first

For example: 30 seconds, 9:16, Chinese, knowledge-based, book introduction, for X / TikTok / Xiaohongshu.

2️⃣ Write the narration script

Let Codex write a 30-second Chinese narration first; don't rush into visuals.

3️⃣ Confirm the visual style

For example: dark editorial knowledge video, high-end knowledge style, dark background, book cover as the main visual, slight camera zoom-in.

4️⃣ Create timestamp storyboards

Break the narration into 4-6 visual segments, each corresponding to a core piece of information.

5️⃣ Prepare materials

Book cover, title, keywords, logo, background music, voiceover text.

6️⃣ Generate voiceover

You can use Edge TTS, ElevenLabs, or Hyperframes media-related workflows.

7️⃣ Transcribe the voiceover

Use the final audio to generate a transcript; don't manually guess subtitle timings.

8️⃣ Create Hyperframes composition

Write index.html using the correct data-start / data-duration / data-track-index.

9️⃣ Create subtitles

Automatically sync subtitles based on the transcript.

🔟 Preview

Watch the preview first; do not render the final video directly.

1️⃣1️⃣ Validate / Inspect

Check for errors, fonts, layout overflows, and subtitle obstructions.

1️⃣2️⃣ Render MP4

Render only after confirming everything is correct.

These are the general steps. In fact, many parts do not require manual intervention, so just understanding these general steps is enough; the manual part is actually quite minimal.

3. Initial Manual Deconstruction + Script Generation

Now we need to deconstruct the video.

First, we need to use Claude Video to deconstruct the video structure. The one I deconstructed was a book account, so its structure is very simple: the first few seconds need to retain the audience, followed by a book introduction or review.

The next steps for narration script generation are as follows:

1. Collect information: (a) Use Codex to collect book information. (b) Call the WeChat Reading Skills connection to see frequently highlighted sections and comments. (c) Have Codex search for famous book reviews.

1. Generate introduction: Let Codex generate an introduction based on book facts. This introduction must "remove the AI flavor" for self-media platforms. You can adjust it based on your aesthetic needs, with specific requirements including: (a) Avoid typical AI expressions like "It's not... but rather...". (b) Avoid sentence structures with too much parallelism. (c) The expression should be smooth and natural, telling the story of the book plainly rather than introducing it in an arrogant tone.

This part requires us to operate based on two points:

1. Based on the previously deconstructed video structure.
2. We need to act as a gatekeeper to review whether the copy is feasible.

Once generated, you can turn the content collected and organized by Codex into a Skill. This way, the narration scripts can be reused in the future.

Everything should be written by Codex based on facts. After this operation, it will give me a narration script, completing the first overall script production.

4. Audio TTS File Processing: Achieving a Natural, Magnetic Voice

After the narration is ready, another major part of the work—which I haven't automated yet by connecting a voice generation API—is audio adjustment.

However, I believe it can be done, so I will write down how to adjust the voice to make it more natural, magnetic, and storytelling-oriented.

I previously shared in a tweet that I used ElevenLabs, which Codex recommended. The English voices in it are truly perfect, with voices from various countries available.

But when using Chinese voice generation, I found it doesn't recognize Chinese very well, resulting in many typos. In this case, it's not just an "AI flavor" issue; it's simply unusable. The solution Codex gave me was to use Jianying or the BytePlus (Volcengine) API first.

Since I don't have the BytePlus API yet, I used the Jianying app on my computer.

My method is very simple:

1. Create a new project and find the text option.
2. Use the "Text to Speech" function and pick a voice I like.
3. Paste the narration copy directly and click generate.
4. Finally, export the voice (Jianying can only export MP4 files).
5. I feed this MP4 file to Codex.

Based on Codex's previous analysis of the voice, I ask it to help me process the sound to be as natural, storytelling, and magnetic as the benchmark video.

Codex's process for handling sound is as follows:

1. Automatic analysis: Codex first analyzes the sound parameters of the original video and modifies my MP4 file accordingly.
2. Parameter settings: I will put a screenshot of the relevant parameters here. You can send the screenshot to Codex, and it will process that effect for you.
3. Simple method: Or use an even simpler method—directly give it a segment of the benchmark video and tell it, "I want the sound processed like this," and Codex will handle it for you.

The sound processed this way is perfect. Once you master this sound processing capability, you can reuse it in other types of video production, such as:

(a) Movie explanations or story narrations

(b) Life philosophy or inspirational videos

(c) Science popularization accounts

In these scenarios, this sound processing method is completely applicable.

As I mentioned, this part should ideally be automated. In self-media creation, avoid manual work whenever possible; in such a "painless" state, it's much easier to persist. So later, I will try using the BytePlus API to let Codex automatically generate narration and voice TTS files.

5. Hyperframes for Visuals, Subtitles, and Audio Alignment

Next is creating the visuals and matching the voice with the subtitles. At this stage, I didn't interfere at all because Codex already had a general plan. I had already fed it a benchmark video, and now with the narration and voiceover, Codex creates the video on its own.

The version generated after video creation won't be perfect the first time; several situations might occur:

1. Voice and subtitles don't match You need to communicate with Codex and let it align them. Codex will automatically extract frames to check why they don't match; this might require about two adjustments.

1. Visual issues Including visual content, subtitle positioning, and the resulting forms. These can be repeatedly communicated with Codex, constantly reminding it to move closer to the video we want to replicate.

Once the entire process is smoothed out, we can generate videos with a high degree of replication. If we can automate the entire workflow, we can run accounts in batches. Whether it's a book account or other types, they can all be deconstructed and replicated following this process.

6. Summary

Finally, I hope this article is useful to everyone.

I am also in a state of learning while doing, and I hope we can communicate and exchange ideas more. Often, it's not that we can't do it, but that we haven't taken action yet.

The same goes for using Codex; as long as we start taking action and chatting with Codex in natural language, we can find these solutions step by step. I believe that as long as everyone starts doing it, everyone can use AI to make a living in self-media.

I hope everyone can use this AI era—an era where ordinary people can infinitely approach high productivity—to do self-media, to create, and to find ways to replicate so that it can generate wealth even while you sleep, moving infinitely closer to freedom.

I am Serena, exploring AI × Web3 × Self-Media, recording how ordinary people take back the initiative in their lives. I hope this article helps you!

👏