Selected Work · AI Content System

From one video
to a production pipeline

I organized topic selection, scripting, visuals, voice, captions, covers, platform copy, and quality checks into a repeatable AI content pipeline. The goal is not merely to finish one video faster, but to make every episode inherit what the previous one taught me.

Role
Product design + content systems + AI collaboration
Stack
React · Remotion · FFmpeg · TTS
Status
Operational · Continuously improving
The Problem

The expensive part of video production is not any single edit. It is re-deciding everything for every episode: what to say, how to structure it, what to show, how audio and captions align, and what is still missing before publication. When a dozen steps depend on memory, finishing one video sends the next one back to zero.

How It Works

From an episode to a complete release

Content is data, components handle expression, and the final voice track owns the timeline. Five stages create a delivery chain that can be checked and reproduced.

🧭
STEP 1
Define the episode
Lock the series, audience, single question, and conclusion before the script expands
🧾
STEP 2
Structure the episode
Put narration, captions, scenes, assets, cover, and platform copy in one configuration
🎙️
STEP 3
Build voice and timing
Generate one continuous narration track and align scenes and captions to real pauses
🎬
STEP 4
Render and verify
Render visuals, captions, and audio, then check dimensions, duration, loudness, and safe areas
📦
STEP 5
Ship one release
Package the video, cover, captions, platform copy, configuration, and QA evidence together

The division of labor is explicit: skills preserve judgment, the episode configuration carries the content, software executes consistently, and a human makes the final call. AI does not decide what I believe; it handles the repetitive work after the decision.

Real Outputs

One foundation, multiple visual languages

These are frames from actual renders, not concept art. Structured content, voice, captions, and QA remain consistent while the visual language changes with the subject.

Four rendered frames showing a problem, a three-layer workflow, a release package, and stable execution
Structured explanationUse flows, comparisons, and conclusion cards to make an abstract method concrete.
A sequence of office-comic scenes showing an urgent task, a model test, a failure check, and a delivery decision
Office-comic micro dramaUse continuous scenes for tasks, conflict, and reversal—not merely a new skin on a slide deck.
The Release

Done means more than an MP4

Everything a platform needs is prepared in one release package, with each item open to inspection and traceability.

9:16
Vertical video
30
Frames per second
2
Platform-specific copy sets
1
Complete release
🎬
Video and cover
Ready for upload
💬
Caption files
SRT + timeline
✍️
Publishing copy
WeChat Channels + Xiaohongshu
✅
QA evidence
Parameters + checklist + manifest
Why It’s Different

Not “one-click viral.” Repeatable judgment.

🧱
Content and components are separate
Each episode lives in structured data; scene components focus on expression, making revision and reuse explicit
🎧
Audio is the source of time
One continuous narration track keeps scenes and captions aligned to natural pauses without jumps in energy
🎨
One foundation, multiple visual modes
Structured explainers, evidence-led layouts, office comics, and generated shots can match the subject instead of converging on one “AI look”
🔎
Every release is inspectable
Video parameters, voice settings, captions, QA, and file manifests remain together so lessons carry forward
Takeaway

From topic to publication, the work no longer breaks into a dozen projects waiting on one another.
The pipeline does not replace creative judgment. It makes each judgment executable, inspectable, and reusable.

More selected work