Hello everyone, I am Chengfeng.
I spent a week using editing skills to create an editing Agent.
It's really satisfying!
It can automatically edit a 30-minute video in 10 minutes.
I often use CapCut (Jianying) to edit talking-head videos, but after using it for a while, I found several problems:
Problem 1: Smart cutting cannot understand semantics
Because it cannot understand semantics, some repetitive segments are not identified. If I speak for twenty or thirty minutes in one go, editing it myself is very tiring.
Problem 2: Subtitle quality is poor
Automatically generated subtitles have many typos.
So I used the Skills feature of Claude Code to make an editing Agent.
The fundamental difference is simple:
Key Differences:
- CapCut = Fixed tool + Manual operation
- Agent = Adaptive system + Automatic learning
I'm not replacing CapCut with a better algorithm, but replacing it with a self-improving system.
But that's not the most impressive part.
The most impressive part is: It understands you better the more you use it, and it gets faster and faster.
Three Core Designs
1. Agent Logic
Only 4 steps are needed:
2. Skills System
Previously, I made all functions into one big skill, which required instructions to distinguish different tasks, which was very troublesome.
Now I have made the 5 core tasks of editing into 5 independent skills and put them in the .claude/skills/ directory. This is clearer and easier to choose.
When you input /v, Claude Code will automatically list these 5 available skills:
You pick one, and the AI executes that skill. Simple, right? What used to take 10 minutes of manual operation now only requires selecting from a menu.
All methodologies are written in the Skills, so you don't have to repeat instructions every time.
The core reason for dividing into multiple independent skills is: humans need to check the phased output. For example, I need to see if the review draft is correct, and then execute the editing after confirmation.
3. Self-updating Mechanism: The more you use it, the better it understands you
This is my proudest design. After each task execution, you can give feedback, and the AI will permanently save the feedback into the Skills.
Key point: Skills gradually change from a set of general rules to a customized solution that belongs only to you.
Use it 10 times, and it knows 80% of your habits. Use it 50 times, and it fully meets your needs.
The more you use it, the more personalized the Skills become, and the better it understands you —— this is the power of self-updating.
Specific Operations
Step 1: Download Skills
Open a new folder; you can use this folder as a dedicated space for processing editing in the future, which is more convenient.
Help me download the editing Skills:
- Address: https://github.com/Ceeon/videocut-skills
- Storage path: .claude/skills/ (Clone directly to this directory, do not create subfolders)
Then it will download on its own, and this process is very fast.
Then we restart to see these skills.
Step 2: Install Environment
Input /v and select videocut:install.
AI will automatically install dependencies and download models (about 5GB):
- FunASR: For slip-of-the-tongue recognition
- Whisper: For subtitle generation
These two models each have their advantages: FunASR is suitable for word-level editing (recognizing slips of the tongue, filler words), and Whisper is better for subtitle quality.
Step 3: Edit Talking-Head Video
Input /v, select videocut:edit talking-head, and then tell the AI your video file path.
AI will automatically:
- Transcribe the video
- Identify slips of the tongue (check sentence by sentence, missing none)
- Identify filler words (um, ah, eh, etc.)
- Identify silence (≥1s)
- Generate a review draft
In the review draft, the stop symbol (⏸️) indicates a silent segment that needs to be deleted. The wavy line after the red dot (🔴～～) indicates a repetitive segment that also needs to be deleted. If you want other formats, you can tell the AI to modify it to a format that meets your requirements.
Key: After the review draft comes out, you look through it. If there is anything you are dissatisfied with (e.g., 'I want to keep more ums' or 'this word was recognized incorrectly'), continue to communicate with the AI until satisfied.
Then input /v, select videocut:self-update, and the AI will record this adjustment into the Skills.
After confirmation, the editing is automatically executed.
Step 4: Add Subtitles
Point: Custom dictionary = Double the subtitle accuracy
A dictionary file is preset in the Skills:
.claude/skills/videocut:subtitles/dictionary.txt
Open this file, and you will see the preset vocabulary:
You only need to add your proper nouns to this file. For example, your company name, product name, channel name, etc.:
Chengfeng's Channel
AI Product Freedom
My Brand Name
Pre-recording into the dictionary allows the AI to automatically use your dictionary to correct model errors when adding subtitles. Accurate the first time.
Then input /v, select videocut:subtitles.
AI will automatically:
- Whisper transcription
- Dictionary correction of typos (using your custom dictionary)
- Generate subtitle files
- AI will let you confirm if the subtitles are correct
- After confirming they are correct, burn them into the video
Key: If you later find that something was missing from the dictionary, you can also use /v to select videocut:self-update to supplement it at any time.
Four steps to get it done,全程自动化.
Summary
Finally, the result is basically consistent with the result of manual editing in CapCut!
If you often edit talking-head videos, I highly recommend trying it. Remember to use the self-updating mechanism, so the Agent will understand you better the more you use it.
