In the past, when making a video, scripting, dubbing, editing, and subtitles had to be learned separately. Just getting started was enough to make people give up. Now, these tasks can be handed over directly to Codex.
You just need to give it an article or a clear theme, and then hand it the corresponding video Skills, and Codex can continue to complete the oral broadcast, sound, Timeline, lens splitting, subtitles, animation and filming. You don't need to learn to edit first, nor do you need to understand the code. As long as you can clearly explain the video you want, you can start.
This article runs the entire process from the beginning: first make a video of the article without showing up, and then introduce the editing method for live oral broadcasts. When you are exposed to a video for the first time, you can also follow it and run out of your own first edition.
The entire process can be simply understood as:
Article or theme → Oral script → Dubbing → Timeline → Splitting → HyperFrames → Check into pieces
Give the video Skills to Codex first
Skill can be understood as a job specification for Codex. It tells Codex what tools to call when encountering oral broadcasting, dubbing, camera splitting, or editing tasks, in what order to complete them, and which problems to check in the end.
You don't need to study a bunch of commands yourself when installing. Directly hand over the corresponding GitHub repository link to Codex and have it read the project description before completing the installation and environment check.
In this process, HyperFrames has provided several different divisions of labor:/faceless-explainer Be responsible for organizing articles or themes into non-camera explanations,/hyperframes-creative Responsible for oral rhythm and camera division,/media-use Manage pictures, sounds and subtitles,/hyperframes-animation Responsible for picture changes. Usually you just tell Codex to "use/hyperframes to make this video" and it can continue to select the parts it needs.
If you need additional visual style, you can add other Skills. For example, if you want to make a collage knowledge video, you can refer to it. HyperFrames Community Skills in vox-explainer; If you want to do the effects of handwriting or drawing, you can use p5-paint-animation; If you want to do product introductions, interface demonstrations and promotional videos, you can refer to video-shotcraft。
These are all extensions, and the first video does not have to be installed all.
Create a new folder, for example codex-video, open it in Codex. Put in the articles, pictures, and audio generated later, and then hand over the HyperFrames repository to Codex to complete the preparations. Next, you just need to confirm the results step by step.
Step 1: Turn the article or topic into an oral script
The oral script is the skeleton of the entire video. No matter how beautiful the picture is, if you don't catch anyone at the beginning, it will be difficult to read the following content.
If you only have one theme, you can ask Codex to write it. For example:
我要做一条 60 秒竖版视频，主题是“第一次用 Codex 做视频”。
观众只会使用 Codex，没有写过口播稿，也没有剪过视频。
开头先告诉他最终能做出什么，再讲清楚：生成口播、选择声音、用 HyperFrames 做画面、检查成片。
每句话尽量能直接念出来，不写书面化小标题，不堆概念。
结尾让他先用一个主题跑通第一条视频。
现在只交付口播稿，暂时不要做分镜和视频。
The more specific the theme, the easier it is to write an oral broadcast. Rather than saying "making AI videos", it is better to directly limit it to "making a tool review into a 60-second vertical version", so that Codex knows what content should be left.
If you already have an article, save the article as article.md, and then change Codex to a version suitable for speaking:
请读取 article.md，把它改成一条 90 秒口播稿。
保留文章的核心判断和最有用的案例，不要逐段缩写全文。
开头先说读者能得到什么，中间只保留三步方法，每一步都用能听懂的句子讲。
文章里的链接、脚注和排版符号不要读出来。
请把最后需要我核对的事实单独列出来。
先只生成 narration.md。
If the article has obvious translation tone, mechanical parallelism, or overly written sentences, you can put humanizer-zh Give it to Codex and let it adjust further without changing facts or personal judgment.
But it's best to read it yourself after you finish correcting it. Many words may seem fine, but when you actually read them, your mouth will tell you what is unnatural.
If you only have the material and haven't thought about the audio broadcast and footage, you can also directly download the video from HyperFrames /faceless-explainer To start, let it organize the explanation directions, narration, and shot splits based on the theme, article, or notes.
No matter which method you choose, it is best to follow a simple structure for oral broadcasting:
Hooks → Problems → Methods → Next step
For example, the beginning of this tutorial can be directly said:
"I don't know what to do with installing Codex? You can use it to make a video first. Give it an article, and it can continue to process oral broadcasts, sounds, subtitles and pictures. Let's run the shortest process first today."
Step 2: Determine the final sound
After the final version is broadcast orally, it will be decided who will read it.
For the first time making a video, ordinary TTS is enough. You can use the existing text-to-speech function, or you can add the open source edge-tts Give it to Codex to complete the installation, select Chinese sounds, and generate audio and subtitles.
It is recommended to generate ten seconds of test audio first, confirm the sound, speed and pause, and then generate the full text. The advantage of ordinary TTS is its speed, which is especially suitable for verifying content and screen flow first.
If you want to use your voice for a long time in the future, you can consider CosyVoice。It can generate a new oral broadcast with similar speaker characteristics based on a reference recording.
Three documents need to be prepared here:
reference.wav: Reference recordings that you or have been authorized;reference.txt: Refer to the complete text actually spoken in the recording;narration.md: New oral broadcast of this video.
The reference text must correspond exactly to the recording. The recording should also be kept as quiet, clear and natural as possible. Background music, echoes and heavy noise reduction will affect the effect.
When installing CosyVoice, hand over the official warehouse directly to Codex:
请帮我安装 CosyVoice：
https://github.com/QwenAudio/CosyVoice
请先阅读官方 README，再检查当前电脑环境并完成安装。
安装好后只生成一条短测试音频，不要直接生成整篇口播。
Listen to the test audio first after installation. When formally generating, don't stuff the entire article in at once. Instead, break it into several paragraphs according to semantics. Listen to each paragraph first to confirm the reading of proper nouns, English and numbers, and then synthesize the final audio.
If you don't need to reproduce the sound for the time being, you can just skip CosyVoice. The latter HyperFrames only requires the final audio and doesn't care which TTS the sound comes from.
Step 3: Let the picture follow the final sound
There should now be two core documents:narration.md and final.wav。
From this moment on,final.wav is the real time ruler for video。
A Timeline is first established based on the final audio, and then subtitles are generated. Oral scripts determine what subtitles are written, and audio transcriptions determine when subtitles appear.
Don't allocate time evenly based on words. There are also ten words, some can be finished in half a second, and some may stop twice in the middle, so subtitles, keywords and shots should follow the real sound.
You can tell Codex directly:
请以 final.wav 为准建立时间轴。
用 narration.md 核对字幕文字，用音频转写结果取得每句话的真实开始和结束时间。
输出 captions.srt 和 timing.json。
如果口播稿与实际声音不一致，请列出差异，不要自行猜测。
The advantage of this is that even if the picture is modified later, it will not be reworked repeatedly due to re-estimating the time.
Step 4: Make a split lens according to the oral broadcast
You don't actually need to learn too much technical terms to split the lens. Cut the oral broadcast into 4 to 8 paragraphs, each paragraph answering only one question:
When the audience hears this sentence, what should they most see on the screen?
Screenshots, photos, charts or role cards that already exist in the article will be used first. Let the audience see the whole thing first, and then zoom in on what is being said. Where there is no material, you can supplement it with keywords, step cards, comparisons, timelines and simple information maps.
Let Codex generate split lenses based on the oral broadcast and Timeline:
请根据 narration.md、timing.json 和 assets 文件夹制作分镜表。
每个镜头写清楚：
对应口播、开始时间、结束时间、主要画面、需要的素材、字幕位置。
原文已有图片必须安排实际展示；
缺少素材的地方先标记，不要用无关图片填满。
If you don't know what to do with kinetic efficiency, you can refer to it. HyperFrames Launches full project in the, or use video-shotcraft The lens formula in the.
The first video should not pursue complex animations every second.First explain the information clearly, and then use one or two simple dynamics to enhance the rhythm.
Step 5: Get HyperFrames to make the first version
Now that you have the oral broadcast, final sound, subtitle time and shot split, you can officially hand it over to HyperFrames:
使用 /hyperframes，把当前项目做成一条 9:16 竖版讲解视频。
请读取 narration.md、final.wav、captions.srt、timing.json、storyboard.md 和 assets 文件夹。
画面跟随真实音频时间，字幕不要挡住主要内容。
优先使用 assets 里的原图和角色卡，缺少素材的镜头用简洁的信息图。
先完成前 15 秒样片并打开预览，我确认方向后再继续全片。
It is specially recommended here to do it first 15-second sampleDon't directly render the entire video.
Looking at these 15 seconds, we will focus on checking three questions: whether the beginning is fast enough, whether the subtitles are clear, and whether the picture really explains the oral broadcast.
If the direction is correct, continue to generate the entire film.
When revising later, don't just say "be more advanced" and "look better". It is best to tell Codex the exact location, such as:
"The title in the sixth second is too small."
"This screenshot stays for at least 3 seconds."
"The subtitles block the characters."
"The first shot entered too slowly."
This kind of feedback will be easier to implement than abstract aesthetic requirements.
After final confirmation, let Codex render:
把当前确认的版本渲染成 final.mp4。
完成后检查文件能否正常播放，并检查声音、字幕和画面是否完整。
The preview page can be opened, which only means that the project can run. The true criterion for completion is final.mp4 It has been generated, and you have watched it from beginning to end to make sure the sound is not truncated, the subtitles are not misaligned, and the pictures are not incorrectly cropped.
How to choose HyperFrames and Remotion?
Both can turn text, sound, pictures and Timeline into videos, but the suitable working methods are different.
HyperFrames It's more like reorganizing the screen based on this issue's content, which is suitable for quick essay videos, knowledge videos and explanatory content.
Remotion It is more suitable to make a set of fixed column templates first, and then fill in the content in the fixed position in each issue.
If you want to take the Remotion route, you can take Remotion Agent Skills Hand it over to Codex, let it complete the installation, and then use /remotion-create Create videos,/remotion-studio Preview,/remotion-render Render.
upstream generated narration.md、final.wav, subtitles, Timeline and pictures can continue to be used without having to do it again.
My advice is:The first one is run with HyperFrames first. After several consecutive episodes, if you find that the lens structure has stabilized, ask Codex to organize it into a fixed column for Remmotion.
If you shoot a live video, add video-use
The previous process can be completely invisible. When you start filming live live audio broadcasts, you can put narration.md Put it next to the camera and shoot in segments. Stop and start again from this sentence. You don't have to start from the beginning every time.
video-use It's more suitable for dealing with live material that's already been filmed. It transcribes the video, finds out what is said in each passage, sorts out repetitions, pauses, misstatements, and restarts, and finally goes into editing.
When installing, let Codex read the repository directly:
请帮我安装 https://github.com/browser-use/video-use。
先阅读 install.md，完成依赖、FFmpeg 和 Codex Skill 注册。
安装后先不要转写任何视频，等我把素材放进文件夹。
After shooting, place the original video into a separate material folder:
请使用 video-use 处理这个文件夹里的真人口播。
先盘点素材并转写，找出重复、停顿、说错和重新开始的位置。
先给我一份剪辑方案，列出准备保留和删除的片段；我确认后再剪。
定稿后添加中文字幕，并在讲到步骤时使用 assets 里的对应图片补充画面。
最后输出预览版，检查无误后再输出 final.mp4。
In this way, the process of real-life live broadcast is:
Oral script → live-action shooting → video-use collating effective clips → subtitles and supplementary pictures → filmmaking
HyperFrames and video-use can also be combined: video-use is responsible for cutting the live broadcast smoothly, while HyperFrames is responsible for information screens such as titles, step cards, screenshot presentations, and titles and ends.
Finally, turn a video into a content system
What is really valuable is not making a video with Codex, but making the same process sustainable.
For example, you can set a theme: tear down an AI tool every week, solve an office problem each time, or explain an industry case each time. The clearer the content object, the easier it is to form a fixed structure behind it.
The whole cycle is actually very simple:
Question from the target user → Select a specific theme → Write a verbal broadcast → Make a video → Read comments and feedback after the broadcast → Decide on the next one
The hook for getting customers should also fall on the specific results.
"I made a video using AI" only proves that you can use tools.
"I turned a 3000-word tool review into a 60-second novice tutorial" to stop people who really need this.
So don't rush to build a complex automation system in the first stage. First choose a topic that you can talk about for ten consecutive issues, and use Codex to run through the first one:
Final version of oral broadcast → sound generation → Timeline alignment → lens splitting → HyperFrames release → inspection
The second article begins to reuse the file structure, and the third article begins to retain fixed title titles, subtitle styles and shot structures.
By doing this, Codex has truly gone from "helping you make a video" to a workflow that can continue to help you produce content.