- Home
- Skills
- Image & Video
- AI-Editor: Raw Footage to Ready-to-Post Reels & Shorts
More screenshots
Works with the AI Agents you already use
AI-Editor: Raw Footage to Ready-to-Post Reels & Shorts
AI-Editor turns Claude Code (or Codex) into a careful video editor for TikTok, Instagram Reels and YouTube Shorts.
$3
AI-Editor: Raw Footage to Ready-to-Post Reels & Shorts
Example session with this skill installed
/video-edit — I put a new recording in raw/: IMG_4821.MOV, about 2 minutes of me talking to the camera about 3 ways to save on groceries. It's in English. Make it ready for Reels.
- Read your context and instructions
- Compiled the ai-editor
Step 1: confirming files (asked once)
I found 1 new file in
raw/:
IMG_4821.MOV (iPhone, 1:58, HDR, landscape). I'll treat it as a
talking-to-camera video in
English, titled
"3 grocery hacks", using your style.md. OK?
Buyer: "yes"
Step 2: suspect words (before any subtitle is burned)
Transcribed locally with Whisper large-v3. Please check these before I add subtitles:
- aldi (00:14): "…I switched to aldi last year…"
- 40 (00:37): "…saves me about 40 dollars a week…"
- costco (01:22): "…a costco membership only pays off if…"
Reply briefly, e.g. "1 → Aldi, 2 ok".
Buyer: "1 → Aldi, 2 ok, 3 → Costco. Remember them."
Corrected, and saved Aldi and Costco to
glossary.mdfor future videos.
Your draft is ready: output/3-grocery-hacks/draft-v1.mp4
| Before | After | |
|---|---|---|
| Length | 1:58 | 1:12 |
| Format | 3840×2160 HDR, landscape | 1080×1920, SDR, 30 fps |
| Voice loudness | −23.4 LUFS | −14.0 LUFS (peak −1.2 dBTP) |
What I cut (46 s): 2 retakes of the intro (kept your last, most fluent one), 11 fillers ("um", "so, uh"), one sentence you abandoned at 0:51, and long pauses shortened to 0.12–0.28 s. Your voice speed is untouched; breaths between sentences are kept.
Picture: HDR converted to SDR, 9:16 window placed on you (framing.x = 0.42), light punch-in on alternate cuts.
Graphics (3, each with a reason):
hook-titleat 0:00, "3 grocery hacks that actually work" (hook)number-calloutat 0:21, "$40 / week" (number)step-labelat 0:09, 0:27, 0:48, "1 / 2 / 3" (list steps)
Quality check: all passed ✅
| Check | Result |
|---|---|
| Specs (1080×1920, H.264, AAC 48 kHz) | ✅ |
| Loudness ±1 LU, true peak | ✅ |
| Leftover silences > 0.45 s | ✅ none |
| Subtitles (every confirmed word shown, safe zone) | ✅ 214 words |
| Black / frozen frames | ✅ none |
| Audio–subtitle sync | ✅ |
I also looked at the QC contact sheets: framing, skin tones and every subtitle frame look right.
One thing to check: at 0:33 there is a jump cut where you turned your head. If it bothers you, I can keep the longer take there.
Everything I decided is in output/3-grocery-hacks/edit-notes.md.
Next steps: watch the draft and tell me what to change in plain words, for example "text a bit smaller", "cut the part at 0:40", or "add quiet music". You'll get draft-v2.mp4; v1 stays as it is.
Connects securely to your tools. The creator never sees your data.
About this skill
You film it. Your AI agent edits it. AI-Editor turns Claude Code (or Codex) into a careful video editor for TikTok, Instagram Reels and YouTube Shorts. Put your raw recording in raw/, say "edit my new video", and get back draft-v1.mp4 at 1080×1920: tight, subtitled, loudness-balanced and checked frame by frame before you ever see it.
No timeline. No keyframes. No 2-hour cleanup per minute of talking. Just a conversation.
The problem it solves
Editing a one-minute talking-head video by hand means scrubbing for every "um", every bad take and every awkward pause, typing or fixing subtitles word by word, then guessing at audio levels. Asking a generic AI to "edit my video" is worse: it improvises FFmpeg commands, forgets to check its work, clips words, and gives you something different every time.
AI-Editor fixes the workflow. The mechanical parts (cutting, subtitles, audio, export, checks) run through a tested command-line toolkit that behaves the same on every machine. Your agent spends its intelligence on the judgment calls: what to cut, which shot goes where, and whether a graphic actually earns its place.
What you get
- Clean cuts that respect your voice. Long pauses, fillers ("um", "uh", "ờ") and retakes are removed, keeping your last good take. Cuts never land inside a word, natural breaths are kept, and loud cut points are checked against the waveform so there are no clicks.
- Word-by-word subtitles that are actually correct. Local Whisper (large-v3) transcription, then the agent asks you about suspect words (names, brands, numbers, unclear audio) before anything is burned in. Answer
1 → Zalo, 2 okand you're done. Corrections can be saved to a glossary, so it learns your names. - Your own style, measured, not guessed. Give 2–3 videos you like. The agent downloads and measures them (cut rhythm, pause length, subtitle size and position, colour, music level, the first 3 seconds) and writes a
style.mdwith exact values, plus a preview image. It takes the rules, never the identity: no copied logos, captions, music or catchphrases. - Graphics with a reason. Five animated templates (hook title, number callout, step label, keyword pop, lower third), used only for a hook, a number, a list step, a key word or a name. Never over a face, never over the subtitles, never "because it looks cool".
- Platform-ready audio. Voice set to −14 LUFS (the level TikTok, Reels and Shorts expect) with a −1 dBTP peak ceiling, gentle compression, noise reduction only when needed, and your own music looped, faded and ducked under speech.
- Automatic quality control. Every render is checked for specs, loudness, leftover silences, missing or overlapping subtitles, black or frozen frames, decode errors and sync. A draft only exists if it passes, and the agent looks at the QC contact sheets itself before showing you.
- Revisions in plain language. "Text too big." "Music quieter." "Cut the part at 0:12." "Wrong word at 0:20, it's Hải Châu." Times you mention are mapped back to the source automatically. You get
draft-v2.mp4; v1 is never overwritten. - A full paper trail. Every decision (what was cut and why, which graphics and why) is written to
edit-notes.mdin your language.
Two formats, many languages
- Talking to camera: your recording, cleaned up and subtitled, with optional cutaways to your clips and photos.
- Voiceover: your own voice recording, or an AI voice from MiniMax reading your script (optional, you bring your own key), laid over your clips and photos, sentence by sentence.
Works in Vietnamese and English (tuned and tested in depth) plus most languages Whisper understands, including Chinese, Japanese, Korean, Thai, Arabic, Hebrew and Hindi, with the right fonts and right-to-left handling built in.
How it works: 4 skills, 4 steps
/video-setup: checks your Mac or PC, shows one list of what's missing with sizes, installs only after you say yes (official sources only: Homebrew, winget, pip, npm), then proves everything works with real tests and a results table./video-style: learns the look of videos you like and writes yourstyle.md. No references? Start from a sensible default./video-edit: transcribe → confirm suspect words → cut → frame → subtitles, music, graphics → render → QC →draft-v1.mp4opens automatically./video-revise: your feedback becomesdraft-v2.mp4,v3… A lasting preference ("always bigger text") can be saved to your style.
You never have to type a command. Talk in your own language; the slash commands are just shortcuts.
Why this beats prompting it yourself
| Prompting an AI from scratch | AI-Editor | |
|---|---|---|
| Consistency | Different commands every run | Same tested toolkit every time |
| Subtitles | Burned in with typos in names and brands | Suspect words confirmed with you first |
| Cuts | Can clip words or leave clicks | Word-safe boundaries, waveform-checked |
| "Make it shorter" | Speeds up your voice | Voice speed is locked; proposes what to cut instead |
| Done means | "It rendered" | QC passed and the frames were inspected |
| Your files | Can be overwritten | raw/ is read-only, every draft is a new version |
Built-in guardrails
Hard rules enforced by the tools, not just the prompt: never change your voice speed or pitch, never burn subtitles before you confirm suspect words, never overwrite a draft or touch raw/, never put reference footage into your video, never install with sudo or from unofficial sources, never print your API key. When a rule blocks something, the agent tells you why and offers another way.
Private by design
Your footage stays on your computer and Whisper transcribes it locally. The only things that leave your machine: what your AI agent sees to do the edit (transcript and frames), and your script if you choose the MiniMax AI voice.
What's inside
- 4 skills (
video-setup,video-style,video-edit,video-revise) with focused reference guides for cutting, subtitles, audio, graphics, voiceover, QC, style analysis and feedback mapping ve.py, a complete command-line toolkit (prepare, transcribe, cuts, render, QC, notes, map-time, style, fonts, doctor…)- 5 HyperFrames graphic templates + a Remotion project
- Bundled open-licence fonts (Be Vietnam Pro, Montserrat, Barlow Condensed, Noto Sans) with on-demand packs for other scripts
- Style and edit-notes templates with validated default values and allowed ranges
AGENTS.md/CLAUDE.mdwiring so Claude Code and Codex both pick it up- A plain-English README written for non-technical creators
Perfect for
- Creators posting talking-head content daily who are tired of manual cleanup
- Coaches, educators and founders turning phone recordings into polished Reels
- Marketers producing faceless voiceover videos from a script and a folder of B-roll
- Agencies that want one consistent, documented editing style across many videos
Output is vertical 1080×1920 only, optimised for videos under ~5 minutes with one main speaker and one main language per video (no bilingual subtitles). Fully tested on macOS (Apple Silicon) with Claude Code; Windows 10/11 and Codex CLI are supported by design but not yet tested. Cutting thresholds were tuned on synthetic voices, so real recordings may need a small style tweak (just tell the agent). Needs ~10 GB free disk; transcription runs locally, so older machines are slower (switch to large-v3-turbo). Login-gated reference links (often TikTok/Instagram) must be downloaded manually. No timeline UI, no AI video generation, no AI music, no auto-posting. Homebrew on Mac must be installed by you. AI voice requires your own paid MiniMax key.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before any payment.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
Fresh listing
Recently published to Agensi
Money-back guarantee
Doesn't do what its page promises? You get your money back.
Trust & safety
Security scan
Security scan passed 9 days ago
Requires
Permissions
Allowed Hosts
Terminal/Shell: runs the bundled ve.py toolkit, FFmpeg, faster-whisper, npm and the installers you approve during setup (never sudo, never curl | bash). Read/Write files: reads your media in raw/ (never modified); writes to work/, output/ (every draft is a new file), references/ (reference videos you ask it to fetch), style.md, glossary.md, and the local .venv / node_modules during setup. Network: one-time setup downloads (Homebrew/winget, pip, npm, pinned font files from raw.githubusercontent.com with checksums, the Whisper model), yt-dlp to fetch reference videos you link, and api.minimax.io only if you opt in to AI voiceover. Environment variables: MINIMAX_API_KEY, read from your local .env and never printed or logged. Transcription runs locally; your footage is not uploaded anywhere by the toolkit.