Nguyen Hong Son · AI Video Editor

I turn raw clips into 400+ Korean Shorts with AI-automated pipelines.

Nguyen Hong Son. Since September 2025 I have designed, automated and run production lines for Korean-market YouTube Shorts: viral ranking compilations, J-POP hook edits, fully AI-generated drama series, 3D cartoon episodes and 2D doodle explainers. One editor, four channels, 10 to 15 finished videos a day.

An editor who builds the machine, then directs it.

I started as a content creator and bilingual MC in Vietnam. In September 2025 I moved into AI-driven video production and built, from zero, the tooling that lets one person research, cut, design, voice, QA and package Korean YouTube Shorts at scale.

My work sits between an editor and a producer: I decide the hook, the rhythm, the on-screen design and the sound, then encode those decisions into prompts, templates and scripts so the next hundred videos keep the same standard. I read the market daily, measure what performs, and rewrite the rules when the data says so.

Short-form craft. 3-second hooks, beat-synced cuts, Shorts-safe layouts, title formulas, A/B titles across channels.

AI direction. Claude Code agents, CapCut JSON automation, Google Flow, Gemini, ElevenLabs, Remotion. I write the rules the AI follows.

Copyright hygiene. Original-creator sourcing, music detection, CC-BY credits, per-video descriptions. Nothing ships without a credit file.

Languages. Vietnamese (native), English (professional), Korean and Japanese market content with AI-assisted localization.

Case 1 · TOP 6 ranking Shorts

From daily trend scan to a packaged Short in under an hour.

The ranking format needs six strong clips, one countdown and a reason to stay to #1. I automated everything that is repetitive and kept my hands on everything that is taste: the pick, the cut point and the layout check.

170videos in 16 days

250+Python tools in the kit

22Kviews, best Short on a 600-sub channel

Trend scanDaily scan of a reference group of Korean ranking channels; topics must be trending within 48 hours and be an event, not a vibe.

Source 8 clips from original creatorsOne clip per creator, no shops, no watermarked reuploads, 720p minimum, screened for text, children and AI fakes. Every clip gets a credit line.

Hook cutMotion and audio-onset detection find the peak; each clip is cut to 4 to 5.5 s ending 1.2 s after the climax, quiet lead-ins sped up, faces reframed so the banner never covers a head.

Build in CapCut from JSONA template project is cloned and only media, text and timing are replaced: header, 6→1 menu, popups, props, foley on the exact impact frame, meme SFX, BGM aligned to the beat.

Safety passMusic detection mutes risky audio, title gets a stroke, speed-ramped clips get foley cover, CC-BY credits are written, Shorts UI safe zone is enforced.

QA and packageMock frames at every popup, menu order check, then export, compress, Korean title and description with per-rank credits, zip handoff and upload log.

Case 2 · AI drama series (Google Flow)

A three-episode revenge drama with no actors, no set and no camera.

Korean "makjang" drama Shorts are a huge niche. I built a channel identity, wrote a serialized script with cliffhangers and produced every shot generatively while keeping the same faces and voices across episodes.

3episodes, 28 to 37 s

100%generated footage and voices

8sper shot, directed second by second

Series writingHottest line first, one twist per episode, "to be continued" card, recap hook on the next part. Titles use the [N편] convention of the niche.

Cast and voicesCharacters generated once and locked, each with a Gemini voice profile, so the mother-in-law sounds the same in episode 3 as in episode 1.

Shot prompts with physicsEvery 8-second shot is scripted per second: one main action, visible contact (coffee actually lands and soaks), no clipping through walls, no new hands.

Batch generation via browser automationPrompts queued and submitted through Claude in Chrome, clips pulled by share link, credits tracked per shot.

Contact-sheet QAOne frame every 0.5 s per clip; any bad physics or face drift means regenerate before assembly.

Assembly and brandWhisper-timed subtitles coloured per character, name tags, stamps, flash and shake on shocks, film grain, channel logo over the AI watermark, one continuous BGM.

Case 3 · 3D cartoon & 2D doodle, rendered in code

When no editor can do it, I write the renderer.

For animated story Shorts I moved off CapCut entirely. The 3D episodes put flat cartoon characters inside a three.js world with ink outlines, depth of field and gliding cameras; the 2D explainers use an SVG character rig I built with two-bone IK limbs, blinking, talking and props, all driven by word-level timestamps from the narration.

8animated episodes, 44 to 173 s

2custom engines (Remotion + three.js, Remotion + SVG rig)

7reusable skills documented for the agent

Character sheet for the zombie apartment series

Reference analysisTimestamped transcript, hook frames at 0.5 s steps, audio behaviour (voice only vs music bed) measured from the reference Short before writing a single line.

Script and castKorean adaptation with new characters (a Joseon magistrate, a zombie-apartment survivor), storyboard per beat, character sheet with poses and expressions.

Narration with timestampsElevenLabs voice generated beat by beat with character-level timing; the timing file drives every animation cue.

Render in ReactScenes authored as React components in Remotion; 3D sets built in three.js with painted textures and ink lines; stills exported for an approval gate before any full render.

Sound designSynthesized SFX placed on visual beats, soft CC-BY music ducked under the voice, loudness-normalized mux, thumbnail baked into frame 0.

I turn raw clips into 400+ Korean Shorts with AI-automated pipelines.

An editor who builds the machine, then directs it.

Four jobs, one person.

AI-automated editing pipelines

Short-form editing & design

Generative video & animation

Research & channel operations

Every format I have shipped, with the videos to prove it.

How a video gets made in my studio.

From daily trend scan to a packaged Short in under an hour.

A three-episode revenge drama with no actors, no set and no camera.

When no editor can do it, I write the renderer.

Tools I use daily, and what I do with them.

Editing & motion

Generative AI

Production & research

Languages

Formats shipped

Also

From the stage to the timeline.

AI Video Editor & Pipeline Builder (independent)

Content Creator

Freelance Bilingual MC & Voice-over Artist

International Student Volunteer & Student Reporter

B.A. International Economics

Need a short-form pipeline, not just a video?