跪拜 Guibai
← All articles
AIGC

9 AI Skills That Automate the Entire Video Production Pipeline

By iThinkAi ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

Video production remains a multi-tool, multi-hour grind for developers who want to publish technical content. A composable, open-source Skill set that chains together inside an existing coding agent turns it into a single-command workflow, with most steps costing nothing beyond the agent runtime.

Summary

Nine open-source AI Skills, installable into Claude Code, Codex, or Cursor, automate video production from idea to final export. The pipeline splits into three groups: mental tasks (topic breakdown, de-AI script rewriting, breath-based subtitles), recording rescue (a director agent that orchestrates sub-skills, automatic filler-word removal, and loudness-normalized audio mixing with ducking), and visual outsourcing (HTML/CSS motion graphics, hand-drawn whiteboard diagrams via Volcano Engine, and templated cover generation).

Each Skill targets a specific pain point. The topic analyzer downloads reference videos watermark-free, transcribes them, and reconstructs their structure into a new script. The editor identifies stumbles and filler words by converting speech to text, then cuts the bad takes and keeps the clean retakes. The B-roll generator finds a visual metaphor for the narration and renders it as a deterministic MP4, avoiding copyright risk entirely.

Most Skills run without paid third-party APIs. Three common pipelines are documented: a full voiceover chain, a viral-video imitation chain, and a text-to-video chain that converts articles into hand-drawn explainers with auto-generated voiceover.

Takeaways
Nine Skills cover the full video pipeline: topic analysis, scriptwriting, subtitles, master control, editing, audio mixing, code-generated B-roll, hand-drawn whiteboard animations, and cover design.
The topic Skill downloads reference videos watermark-free via yt-dlp, generates timestamped transcripts, and reconstructs the structure into a new script.
The script Skill rewrites AI-generated text into spoken language and outputs a shot-by-shot script accurate to the second.
The subtitle Skill segments by breathing and semantic pauses rather than punctuation, and proofreads common homophone errors, outputting SRT/ASS.
A director Skill (`space-video`) interviews the user about format and references, then dispatches to the correct sub-Skills automatically.
The editing Skill converts speech to text, identifies filler words and stumbles, and applies a "delete earlier, keep later" rule to stitch clean takes.
The audio Skill generates TTS via edge-tts, fetches royalty-free background music, applies sidechain ducking, and normalizes loudness to -14 LUFS without re-encoding the video.
B-roll is generated as HTML/CSS motion graphics rendered to deterministic MP4, avoiding copyright issues entirely.
The hand-drawn whiteboard Skill decomposes an article into 6–10 frames of black-line-on-white sketches with a blue accent, rendered via Volcano Engine Seedance.
The cover Skill locks a color palette, font, and layout, swapping only the title and accent color per video, and exports multiple platform sizes in one click.
Installation is a single command: `github.com/SpaceZephyr/creator-buddy/tree/main/video-skills`.
Three documented pipelines cover full voiceover production, viral-video imitation, and text-to-video conversion.
Most Skills require no paid third-party APIs; only video downloading and model generation need separate keys.
Conclusions

Treating video editing as a text problem — transcribe, diff the transcript, keep the good takes — is a pragmatic shortcut that sidesteps the complexity of timeline-based NLEs.

Rendering B-roll as deterministic HTML/CSS animations rather than sourcing stock footage eliminates copyright risk and makes the output reproducible, which matters for technical content where precision counts.

The "director" meta-Skill is effectively a lightweight router that gathers requirements before dispatching, solving the discoverability problem that emerges once a Skill collection grows beyond a handful.

Normalizing audio to -14 LUFS without re-encoding the video is a production detail that most hobbyist tools skip, yet it is what makes the output sound consistent across platforms.

Concepts & terms
LUFS
Loudness Units relative to Full Scale — a standard for measuring perceived audio loudness. -14 LUFS is a common target for streaming platforms like YouTube and Spotify, ensuring consistent volume across different videos.
Audio ducking
A mixing technique where the volume of background audio automatically lowers when a foreground sound (like speech) is detected, then rises again during silence. Also called sidechain compression.
edge-tts
A free, open-source Python module that uses Microsoft Edge's online text-to-speech service to generate natural-sounding speech without an API key.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗