# The AI Video Editor Handbook > Source: https://statur.co/ai-video-editor · Version 1.1, last updated 2026-11-08 · Author: Dr. Louis Korczowski, short-form content strategist · Companion AI skills: https://github.com/lkorczowski/content-strategist-AI-cookbook Handbook · by Dr. Louis Korczowski # The AI Video Editor: an AI agent edits your talking-head videos at one of four production levels, up to motion design and sound design, and refuses what should not ship. The layers of an edit that look the most expensive carry the least evidence. In a TikTok study with 4,477 viewers, high production quality was among the weakest correlates of perceived authenticity, while Meta found that brand ads with speech and music were twice as likely to rank in the top fifth for brand interest. This handbook gives an AI agent a production dial for your own recordings, from the clipping edit to color, sound design, captures of your own pages and motion design, with a quality gate that can refuse a clip. Every recipe ends with one thing to do today and a prompt for your assistant. TL;DR Give an AI agent your talking-head recording and a production level from 1 to 4. Level 1 is the clipping edit: cuts on word timestamps, word captions, a hook box, punch-ins and a master at -14 LUFS. Level 2 adds a clean picture, a voice chain, caption emphasis, a music bed and an effect on each graphic entrance. Level 3 adds cards that prove each claim: captures of your own pages, lists from your own data, a source line under every number. Level 4 adds background replacement, words behind you and motion matched to a video you like. In the ad studies I found, the craft that steers attention (a hook with sound and picture in the first two seconds, a voice with music, your face to camera) paid wherever it was measured, and polish paid in a few places, so the dial adds craft first and polish last. Every insert proves a sentence, and a quality gate measures loudness, safe areas and the words heard back, then refuses what should not ship. The agent does the hours with free tools. You choose the level, review on your phone and post. The demo ## One selfie, edited at level 4 by an agent. I recorded 97 seconds on my iPhone, in flat light, and gave the file to an agent running the skills of this handbook. Here is the edit it delivered, and the same moment before and after. Fig. 1 · The edit, level 4, 91 s● sound on Version 3 of the edit, with sound. The watermark and the end card point to this page. [Open the video file (MP4)](/media/ai-video-editor/demo.mp4). Fig. 2 · Before and after● the same moment ![Before: a frame of the raw iPhone selfie, with the phone's green-screen effect putting a screenshot of my own text behind me.](/media/ai-video-editor/before.jpg)Before ![After: the same moment at level 4, on an ink grid with the 400M+ counter and its source line behind me, captions and the watermark.](/media/ai-video-editor/after.jpg)After Left, a frame of the recording as the phone saved it. Right, the same moment in the edit: SDR plate, grade, keyed background, the counter behind me with its source line, captions. What viewers reward ## The layers that look expensive carry the least evidence. The measured studies on production value point one way. The craft that steers attention (a hook with sound and picture, a voice with music, a face to camera) paid in every study I found. Polish paid in a few. That is why the dial adds craft first. 01 ### A polished look does little in the feed. MAGNA and TikTok showed 59 ads to 4,477 users in the US and the UK, in a controlled version of TikTok with randomized exposure: high production quality was among the weakest correlates of perceived authenticity, and native ads seen as inauthentic were 19% less effective at driving purchase intent ([MAGNA x TikTok, 2022](https://magnaglobal.com/wp-content/uploads/2022/08/MAGNA-TikTok-Understanding-Strengths-of-TikTok-Ads-003.pdf)). TikTok's advice to advertisers: "go for a DIY or not overly polished style" ([TikTok Ads Help Center](https://ads.tiktok.com/resources/help/article/creative-best-practices)). 02 ### Edit-layer craft pays. Meta and Toluna coded 65 variables across 100 Instagram Reels ads, each shown to 100 people in Germany, the UK, Italy and Turkey. Brand ads with speech and music together were 2.0 times more likely to rank in the top fifth for brand interest, and a hook with sound and picture in the first one to two seconds made an ad 1.5 times more likely to rank there for purchase intent ([PPC Land, on the Meta and Toluna study](https://ppc.land/meta-and-toluna-study-finds-emojis-lift-reels-purchase-intent-2-5x/)). 03 ### Your face is the asset. In a Vidmob and TikTok analysis of 1,678 paid TikTok ads, talent speaking direct to camera, talking-head style, lifted the 2-second view-through rate by 14% ([Advanced Television, April 2024](https://www.advanced-television.com/2024/04/03/report-reveals-tiktok-brand-engagement-strategies)). Across 6.9 million viewing sessions on edX, an informally filmed office course drew about two to three times the engagement of a course filmed in a TV studio on videos longer than six minutes ([Guo, Kim and Rubin, 2014](https://www.cs.rochester.edu/hci/pubs/pdfs/edX-MOOC-video-production-and-engagement_LAS-2014.pdf)). Every level keeps your face on screen most of the time and comes back to it after each card. 04 ### Polish pays in a few places. In LinkedIn's analysis of more than 13,000 B2B video ads, cinematic, narrative-driven brand videos showed +129% engagement and short vertical "real talk" videos +103% dwell time, against baselines the coverage does not give ([Search Engine Land](https://searchengineland.com/linkedin-study-b2b-video-ads-engagement-lift-458873)). Explainers whose pictures carry data, maps or product screens are the other case. Level 4 is for those posts. 05 ### Most editing hours go to the layers on top. Ali Abdaal's editors put a typical long-form edit at about 30 hours: 4 of A-roll cleanup, 8 of B-roll, 6 of titles and motion graphics, 4 of music and sound design, and 8 of reviews, refining, sponsor segments and the rest. AI already brings their rough cut under an hour ([Adobe Blog, September 2024](https://blog.adobe.com/en/publish/2024/09/05/how-ai-helps-ali-abdaal-save-hours-every-week), partner content). B-roll, motion and sound make 18 of the 30 hours (calc.), the three layers the companion skills cover. **How far this evidence goes.** The ad studies above measure paid ads, mostly for consumer brands, and each was run with or for a platform. I found no published randomized test of organic founder videos by production level, a music level under a voice, or a cut rate that holds attention in short videos. So the dial ships with a test: render two levels of the same clip and let your own account decide (recipe 01). The method ## One cut, one dial with four levels, one gate. Level 1 is the clipping edit and stays the default. Each level adds layers to the one below. Nine recipes follow: each one fits on a phone screen, ends with one thing to do today, and comes with a prompt for your AI assistant. Fig. 3 · The production dial● the default: level 1 Level What the agent adds Fits Agent time A human editor (US vendor rates) 1 · Basic, the default The clipping edit: cuts on word timestamps, word captions with the active word highlighted, a hook box, punch-ins, a master at -14 LUFS Lives, organic clips, opinion posts, volume tests 72 s for a 1:28 clip on a 4-core virtual machine (measured on the [clipping](/clipping/) run) $15 to $30 for a basic marketplace edit 2 · Native plus A clean SDR picture and a light grade, a voice chain, caption emphasis, a music bed under the voice, an effect on each graphic entrance, one or two captures of your own pages One-message posts, paid feed tests not timed yet $100 to $300 per clip with captions, transitions and basic motion graphics 3 · Proof and explainer Full-screen cards that prove each claim (captures, lists, stats with their source line, diagrams), a name card, B-roll of your own footage Founder posts with numbers, product proof, B2B explainers not timed yet $300 to $1,500 per performance ad 4 · Signature Background replacement, words behind you, moves of the speaker, motion design matched to a reference video Launches, flagship explainers, brand stories about 35 to 45 min for 91 s on 2 cloud vCPUs (calc. from the measured run) $120 to $180 an hour at a US agency Switch, off Synthetic media: generated B-roll, avatars, cloned voices. When it is on, the post carries a label on every platform and the edit keeps the provenance of each generated asset. **How to read the prices.** The human rates come from agencies that sell editing, so read them as vendor data: [beCreatives](https://becreatives.co/video-editing-prices/) for marketplace edits, [ShortVids](https://shortvids.co/video-editing-cost-breakdown-usa/) for the rest. Upwork counts 1 to 4 editing hours per finished minute, at median rates of $10, $40 and $100 an hour for entry, mid-level and senior editors ([Upwork](https://www.upwork.com/hire/video-editors/cost/)). Automated clipping tools charge about $0.10 to $0.25 per source minute (calc. from [OpusClip](https://www.castmagic.io/blog/opus-clip-pricing), [Submagic](https://www.hooked.so/compare/submagic-pricing) and [Captions](https://www.toolsurf.com/captions-pricing-2026/) pricing snapshots). The tools in this handbook are free: the cost is your compute time and your review. When What Time Once Intake, reference video, brand tokens about 10 min Every phone recording SDR plate and grade about 80 s per minute of footage for the plate Every clip Level, cut, planning sheet, render, gate one session, then machine time Every version Contact sheet and phone review a few minutes 1. [01Choose the level for each clip](#r01)_5 min · every clip_ 2. [02Answer five questions once](#r02)_10 min once · when your style changes_ 3. [03Fix the picture before anything else](#r03)_about 80 s per minute for the plate · every phone video_ 4. [04Cut once, then plan one insert per sentence](#r04)_one session with your AI · every clip_ 5. [05Show your own proof on screen](#r05)_one shot list per clip · level 2 and up_ 6. [06Draw the graphics in your brand](#r06)_7 to 16 frames a second · level 2 and up_ 7. [07Build the sound around your voice](#r07)_about 1 min of machine time · level 2 and up_ 8. [08Replace the background and put words behind you](#r08)_14 min of matting for 97 s · level 4_ 9. [09Let the gate refuse a clip, then review and post](#r09)_1 to 2 min of checks · every version_ The recipes ## Nine recipes, from the level to the post. 01 ### Choose the level for each clip 5 min · every clip Set a ceiling once, then let the agent pick the lowest level that proves what the clip says. Steps 1. Set your ceiling: 1 for a clipping batch, 2 for a single video, 3 or 4 for a launch or a flagship explainer. 2. Give the agent the transcript. It marks every sentence that carries a number, a product claim, a place or a named step. 3. It picks the level under your ceiling: a number earns a stat card, a product claim earns a capture of your page, a pure opinion stays at level 1 or 2. 4. With no data on your account yet, ask for a level 1 and a level 3 version of the same clip, post both, and compare. 5. Keep synthetic media off. If you ever switch it on, label the post on every platform. Two separate dials hide in "production value". Attention control (hook, captions, sound locked to picture, proof inserts) carries the evidence: the ad studies reward a hook with sound and picture, a voice with music and a face to camera. Polish (looks, keyed backgrounds, cinematic motion) pays in a few places: B2B brand stories and explainers whose pictures carry data. The studies cover paid ads, so your own account settles the rest. Avoid - Level 4 on every clip - Polish with no claim to prove - A level chosen before reading the transcript TodayPick one recent video and write the level you would give it, with the sentence that justifies it. For your AI Copy this prompt into your assistant: Here is the transcript of my video \[file or paste\]. My ceiling is level \[1 to 4\]. Mark every sentence that carries a number, a product claim, a place or a named step, propose the lowest production level that proves them, and list what that level adds on screen. If my account has no data yet, plan a level 1 and a level 3 version of the same clip as a test. Skill: [statur-video-editor-pro](https://github.com/lkorczowski/content-strategist-AI-cookbook/tree/main/skills/statur-video-editor-pro) · install: `npx skills add lkorczowski/content-strategist-AI-cookbook --skill statur-video-editor-pro` 02 ### Answer five questions once 10 min once · when your style changes The agent asks five things in one message and uses a stated default for every answer you skip. Steps 1. Level ceiling. Default: 2 for a single video, 1 for a clipping batch. 2. A video whose editing you like. Give the file when you have the right to use it: the agent measures its pace, face share and loudness, reads its captions, inserts and color off a contact sheet, and writes all of it into a style card. With a link only, it fills the card by watching and calls it an estimate. 3. What it may show on screen: your site, docs, repository, dashboards and past posts. Third-party pages only with a reason. 4. Music: your licensed track, the platform's library at posting time, or a generated bed (the default). 5. Platforms, caption language and deadline. A reference video turns taste into numbers an agent can match: cuts per 10 seconds, the share of frames where your face fills the picture, caption size and position, inserts per minute. The agent copies the method of the reference; its frames, music and graphics stay with their owner. Avoid - Answering the same questions for every clip - Copying the reference's assets - A reference file you have no right to use TodaySave one video whose editing you would like on yours, and write down what you like in it. For your AI Copy this prompt into your assistant: Before you edit, ask me the five intake questions of statur-video-editor-pro in one message (level ceiling, reference video, what you may show, music, platforms and deadline), with your default for each. Then study my reference \[file or link\], fill the style card (pace, face share, hook layout, captions, inserts, motion vocabulary, sound, color) and tell me which level it matches and what to copy or skip for my brand, using my context file \[file\]. Skill: [statur-video-editor-pro](https://github.com/lkorczowski/content-strategist-AI-cookbook/tree/main/skills/statur-video-editor-pro) · install: `npx skills add lkorczowski/content-strategist-AI-cookbook --skill statur-video-editor-pro` 03 ### Fix the picture before anything else about 80 s per minute for the plate · every phone video Phone footage in HDR becomes one clean SDR picture first, and every caption, card and capture is designed on that picture. Steps 1. Check the file. An iPhone often records HEVC in HLG with BT.2020 colors: the agent reads it with ffprobe. 2. Make the SDR plate: a tone map to BT.709 with a known curve (reference white at 203 nits). The original stays untouched. 3. Measure three frames: levels at 2, 50 and 98%, clipped pixels, and the skin hue against the skin-tone line of a vectorscope. 4. Grade with a small LUT whose settings are plain numbers: exposure, white balance, contrast, saturation with skin protection. Then a light sharpening. 5. Measure again after the grade and look at one still before the full render. Meta's video engineers list white overlay text that looks gray next to HDR video among the challenges of HDR on Reels ([Engineering at Meta, July 2023](https://engineering.fb.com/2023/07/17/video-engineering/hdr-video-reels-meta/)). My test selfie was HEVC in HLG: converted without a tone map, the face looked burnt and the overlays gray. The plate costs about 80 seconds per minute of footage on 2 vCPUs, so it runs first, at every level. Avoid - Graphics designed on the HDR file - Skin smoothing or face reshaping - LUT packs whose license is unclear TodayCheck whether your phone records video in HDR, and note the setting. For your AI Copy this prompt into your assistant: Read \[video file\] with ffprobe and tell me its color transfer and primaries. If it is HDR (HLG or PQ), make an SDR BT.709 plate with a tone map (reference white 203 nits) and keep the original untouched. Then measure levels, clipped pixels and skin hue on three frames and propose a grade as a LUT with its settings in plain numbers. Skill: [statur-color-grade](https://github.com/lkorczowski/content-strategist-AI-cookbook/tree/main/skills/statur-color-grade) · install: `npx skills add lkorczowski/content-strategist-AI-cookbook --skill statur-color-grade` 04 ### Cut once, then plan one insert per sentence one session with your AI · every clip The cut comes from the clipping edit, and every graphic, capture or sound is written next to the sentence it serves. Steps 1. Transcribe and cut with the clipping skills: word timestamps, silences and false starts removed, caption spelling fixed. One cut logic for every level. 2. Look at a frame grid with the caption zone drawn, and put captions where the face is not. 3. Write the planning sheet: one row per sentence, with what the viewer sees, what it proves and the sound. 4. Hook: picture, sound and text within the first one to two seconds. The text hook makes the same promise as the spoken one, in other words. 5. Come back to the face within about four seconds after a full-screen card, and never stack more than two cards in a row. 6. Write the pro spec with every time on the source timeline, so the graphics stay on their words when the cut changes. Source time (s) Sentence The viewer sees It proves Sound 0.3 to 3.2 Personal brand won't be the same after this hack me over a wall of my own clips, the text hook the account posts many clips impact, whoosh on the hook 3.3 to 7.5 I'm posting more than 10 times a day a full-screen list of the day's posts from my posting log the number I said one tick per row The first two rows of the planning sheet of my test, condensed. In Meta's study with Toluna, a hook with sound and picture in the first one to two seconds made a Reels ad 1.5 times more likely to rank in the top fifth for purchase intent ([PPC Land](https://ppc.land/meta-and-toluna-study-finds-emojis-lift-reels-purchase-intent-2-5x/)). A meta-analysis of 57 studies found that showing key terms beats repeating the spoken words verbatim ([Adesope and Nesbit, 2012](https://rex.libraries.wsu.edu/esploro/outputs/journalArticle/Verbal-Redundancy-in-Multimedia-Learning-Environments/99900601157101842)), which is why the text hook rephrases the spoken one. Avoid - B-roll with no sentence to serve - A text hook that repeats the spoken words - Graphics timed on the output timeline TodayWrite the planning sheet of your next clip: one row per sentence, what the viewer sees. For your AI Copy this prompt into your assistant: Here is the words file of my clip \[file\]. Cut it with statur-clip-edit --plan-only, then write the planning sheet: one row per sentence with source times, what the viewer sees, what it proves and the sound. Rules: picture, sound and text within the first 2 seconds, the text hook rephrases the spoken promise, back to my face within 4 seconds after a card, every number with its source line. Skill: [statur-video-editor-pro](https://github.com/lkorczowski/content-strategist-AI-cookbook/tree/main/skills/statur-video-editor-pro) · install: `npx skills add lkorczowski/content-strategist-AI-cookbook --skill statur-video-editor-pro` 05 ### Show your own proof on screen one shot list per clip · level 2 and up Your own footage and your own pages come first. Stock never proves anything, and generated media waits behind its switch. Steps 1. For each sentence, ask what would prove it. A number needs its source on screen, a product claim needs the product, an opinion needs nothing. 2. Look in this order: your past recordings and clips, your own pages (site, docs, repository, dashboards, posts), licensed stock, and generated media only with the switch on. 3. Capture public pages at 2x. For a page only your browser can see, save a self-contained snapshot in your own browser and let the agent render it offline, with private parts hidden or blurred. 4. Cut a few seconds of B-roll from a long recording where the file lives, and move only that piece. 5. Log every asset with its source, owner, license and the sentence it proves. In the same Meta and Toluna study, direct-response ads that showed the product with added context were 5.3 times more likely to rank in the top fifth for purchase intent ([PPC Land](https://ppc.land/meta-and-toluna-study-finds-emojis-lift-reels-purchase-intent-2-5x/)). On my test, captures of my own GitHub repository carried the sentence about the cookbook, and a few seconds of my own live carried the sentence about clipping a webinar. Avoid - Stock footage used as proof - Third-party pages with no reason - Emails or customer names left visible in a capture TodayList three pages of yours that prove what you say most often. For your AI Copy this prompt into your assistant: For each sentence of my planning sheet \[paste\], tell me what would prove it on screen and where it lives among my assets \[site, docs, repository, dashboards, past posts and clips\]. Capture my public pages at 2x, tell me which pages need a snapshot from my own browser, hide or blur any personal data, and log every asset with its source, owner, license and the sentence it proves. Skill: [statur-broll-find](https://github.com/lkorczowski/content-strategist-AI-cookbook/tree/main/skills/statur-broll-find) · install: `npx skills add lkorczowski/content-strategist-AI-cookbook --skill statur-broll-find` 06 ### Draw the graphics in your brand 7 to 16 frames a second · level 2 and up Every graphic enters on the word that calls it, in your colors and fonts, and stays clear of the platform buttons. Steps 1. Turn your visual system into tokens once: colors, fonts, corner radius, and what the brand never does (gradients, shadows). 2. Pick the templates: hook box, word captions with emphasis, chips, name card, list card, page capture with zoom and highlight, prompt box, chart, loop diagram, end card, moving watermark. 3. Time each one on its word. A card enters when its claim is said, a number counts while it is spoken, and a counter never shows zero before the number. 4. Keep text inside the area TikTok and Reels both leave free on a 1080x1920 frame: from y 269 to y 1248, between x 120 and x 960 above y 840, and between x 120 and x 780 below it. 5. Render a few stills first and look at each one before the full render. The templates are HTML pages drawn frame by frame by a headless browser, with a transparent background. The same frame always renders the same picture, so a change re-renders only the frames it touches. The safe area intersects two vendor readings of the platform specs ([Billo on Meta](https://billo.app/blog/meta-ads-safe-zones), [AdKit on TikTok](https://adkit.so/tools/safe-zones/tiktok)); the research found no official numbers for YouTube Shorts. Avoid - A brand font that silently falls back to Arial - A card hidden behind its own panel - Text under the platform buttons TodayWrite your five brand tokens: ink, paper, accent, display font, body font. For your AI Copy this prompt into your assistant: Read my brand context file \[file\] and write my motion tokens (ink, paper, accent, muted, line, display and body fonts, radius, what the brand never does). Then, for my planning sheet \[paste\], pick a template for each graphic, time it on the word that calls it, keep text inside the cross-platform safe area of a 1080x1920 frame, and render 5 stills for me to check before the full render. Skill: [statur-motion-design](https://github.com/lkorczowski/content-strategist-AI-cookbook/tree/main/skills/statur-motion-design) · install: `npx skills add lkorczowski/content-strategist-AI-cookbook --skill statur-motion-design` 07 ### Build the sound around your voice about 1 min of machine time · level 2 and up The voice leads. Effects mark visual events, the music sits well under the voice, and the master lands at -14 LUFS. Steps 1. Clean the voice: high-pass filter, denoise, EQ, de-esser, light compression. Never change what was said. 2. Place one effect per visual event: a card enters, a word lands, a number stops. Ticks and keys stay at least 20 dB under the voice. 3. Lay the music: your licensed track, the platform's library when you post, or a generated bed with no license to track. Keep it about 16 to 18 dB under the voice while you speak. 4. Normalize the master in two passes to -14 LUFS, with a true peak at -1.5 dBTP. 5. Read the mix report, then transcribe the master again: fewer words heard back means the mix covers the voice. A tone synchronized with a visual change makes the target pop out and captures attention ([Van der Burg et al., 2008](https://doi.org/10.1037/0096-1523.34.5.1053)), which is the case for locking every effect to the picture. No study gives a music level under a voice, so the 16 to 18 dB default is a setting you tune and check in the mix report. No platform publishes a loudness target either: Frame.io's guide says YouTube tends to aim for -12 to -14 LUFS ([Frame.io](https://workflow.frame.io/guide/loudness-for-youtube)), and -14 LUFS is the usual convention for short videos. Avoid - An effect over a key word - Music from a library that bans automated downloads - Turning the voice up when the music is too loud TodayWatch your last video on a phone speaker and note where the music covers a word. For your AI Copy this prompt into your assistant: Mix the sound of my edit \[plan and cue list\]: voice chain (high-pass, denoise, EQ, de-esser, compression), one effect per visual event with ticks and keys at least 20 dB under the voice, music \[my licensed track, or a generated bed\] about 16 to 18 dB under the voice while I speak, master at -14 LUFS and -1.5 dBTP. Give me the mix report with the loudness of each stem. Skill: [statur-sound-design](https://github.com/lkorczowski/content-strategist-AI-cookbook/tree/main/skills/statur-sound-design) · install: `npx skills add lkorczowski/content-strategist-AI-cookbook --skill statur-sound-design` 08 ### Replace the background and put words behind you 14 min of matting for 97 s · level 4 A person matte lets the agent swap your background and set big words behind you. The speaker stays real. Steps 1. Run the person matte. Today the skill mattes every frame of the recording (14 minutes for 97 seconds on 2 vCPUs); matting only the ranges an effect needs is on its roadmap. 2. Replace the background on the keyed ranges with a designed one: a wall of your own posts, an ink grid in your colors. 3. For words behind you, scale yourself to about 0.82 of the frame, anchored at the bottom, and set short words above your head. 4. Feather the edge of the plate so shoulders and hands never end on a hard line. 5. Check stills of the hair edges: no halo, no floating speaker. On my test, words behind me on the real room were unreadable: my head covered them. On a keyed ink stage with me scaled to 0.82, short words such as IDEA, VISION and VIRAL read at a glance. The best CPU matte for video I found, [Robust Video Matting](https://github.com/PeterL1n/RobustVideoMatting), is licensed GPL-3.0, so the skill runs it as a separate tool. It took 860 seconds for 2,915 frames on 2 vCPUs. Avoid - Long words behind a close face - Skin smoothing or gaze correction - Matting the whole video when two ranges need it TodayPick the three words of your next video that deserve to sit behind you. For your AI Copy this prompt into your assistant: In my pro spec \[file\], add keyed ranges for \[time ranges\]: run the person matte, replace the background there with \[a wall of my own posts, or an ink grid in my tokens\], scale me to about 0.82 anchored at the bottom, put the words \[words\] above my head, feather the plate edge, then render stills of the hair edges for me to check. Skill: [statur-video-editor-pro](https://github.com/lkorczowski/content-strategist-AI-cookbook/tree/main/skills/statur-video-editor-pro) · install: `npx skills add lkorczowski/content-strategist-AI-cookbook --skill statur-video-editor-pro` 09 ### Let the gate refuse a clip, then review and post 1 to 2 min of checks · every version An edit ships after a measured check, and a clip that fails goes back with its reasons. Steps 1. Run the gate. Hard checks: H.264 with BT.709 tags at the planned size, loudness between -15 and -13 LUFS with a true peak at -1 dBTP or lower, no black stretch, at least 90% of the planned words heard back, an audio track. 2. Read the notes: a duration off the plan, frozen stretches, graphics in a platform button zone. 3. Look at the contact sheet frame by frame: the hook readable at a glance, captions clear of the mouth, every card fully visible, numbers matching their source line, the end card address live. 4. Three decisions always come to you: realistic synthetic people or scenes, captures of third-party pages, and music whose license the asset log cannot prove. 5. A REFUSE goes back with its reasons: fix and render again, or drop the clip. 6. Review on your phone and post it yourself. Keep the spec, the asset log, the mix report and the quality report next to the master. An editor's most valuable output is often the cut that never ships, and an AI judge can wave everything through. In a test reported by CineMontage, the Editors Guild magazine, AI story analysis advised greenlighting every script; a story analyst summed it up: "AI never passed on anything" ([CineMontage](https://cinemontage.org/?p=62166)). A gate with written reasons keeps that refusal in the pipeline. Avoid - Shipping a REFUSE because it looks fine - Letting the agent post - Deleting older versions TodayWrite three reasons you would refuse to post one of your own clips. For your AI Copy this prompt into your assistant: Run the quality gate of statur-video-editor-pro on \[master file\] with my plan and words file: the hard checks (format and BT.709, loudness -15 to -13 LUFS with true peak at -1 dBTP or lower, black stretches, words heard back at 90% or more, audio track), the notes, and a 16-frame contact sheet you read frame by frame. Answer PASS, PASS\_WITH\_NOTES or REFUSE with the reasons, and list the decisions that need me. Skill: [statur-video-editor-pro](https://github.com/lkorczowski/content-strategist-AI-cookbook/tree/main/skills/statur-video-editor-pro) · install: `npx skills add lkorczowski/content-strategist-AI-cookbook --skill statur-video-editor-pro` The receipts ## What the agent measured on my run. My own video, recorded for this test and edited on October 28, 2026, on 2 cloud vCPUs. Every number below comes from the reports the skills write next to the master. Source A 97-second iPhone selfie in HDR (HEVC, HLG, BT.2020), 1080x1920 at 30 fps, flat light, recorded with the phone's green-screen effect Cut The clipping cut: silences and two false starts removed, caption spelling fixed. The master runs 91 seconds with its end card Picture SDR plate first, then a grade as a LUT. Person matte on 2,915 frames in 860 seconds (3.4 frames per second) On screen A hook box in the first three seconds, backgrounds made of my own posts, a card listing the day's posts from my posting log, my name card, a 400M+ counter with its source line (organic views for clients, last count January 2026), a chart, captures of my GitHub repository, B-roll from my own live, a prompt box, a loop diagram, big words on an ink stage behind me, an end card with this page's address Graphics HTML templates drawn at 7 frames per second for the full backgrounds and 12 to 16 for the transparent layers Sound 124 cues locked to visual events, a generated music bed about 16 dB under the voice, a mix at -14.0 LUFS and -1.5 dBTP Compute About 35 to 45 minutes for the first full pass on 2 cloud vCPUs (calc. from the measured steps), plus about 10 minutes lost when one long render died at frame 1,200. The picture now renders in parts of 300 frames that resume where they stopped Quality gate PASS on version 1. On version 3, PASS with one note ("speech check not run"), then the speech check: 95.6% of the caption words heard back (the misses were transcription errors on words I had corrected), -14.0 LUFS with a true peak of -1.4 dBTP, no black or frozen frame, and five safe-area flags on full-screen cards that cover the platform zones on purpose Versions 2 and 3 Version 2: the counter starts on the spoken number, the big words moved to keyed ink stages with me scaled to 0.82, captions move up while I am scaled down, and the edges are feathered. Version 3: my face stays on screen in a window under four of the full-screen cards Limits ## What the editor does badly today. 01 ### Compute Level 4 is slow on a small machine: about 35 to 45 minutes for 91 seconds on 2 cloud vCPUs, most of it in the person matte and the compositing. I have not timed a Mac yet. Render stills first. The matte runs on every frame today, and matting only the ranges an effect needs is on the roadmap of the skill. 02 ### The hook check is manual The gate measures loudness, safe areas and the words heard back. It does not yet score whether the text hook and the spoken hook make one promise in different words: read the hook frame yourself. 03 ### Captures from a sandbox A cloud sandbox can be blocked from your own site and from stock libraries. The skill then takes the capture in your own browser as a self-contained snapshot and renders it offline. 04 ### No hand-off to a human editor yet The edit lives in a JSON spec with its layers and receipts. An export to Premiere Pro or DaVinci Resolve through OpenTimelineIO is planned and not built. 05 ### Licenses The matte runs as a separate tool because [Robust Video Matting](https://github.com/PeterL1n/RobustVideoMatting) is GPL-3.0. Remotion asks for a paid company license above three employees ([Remotion license](https://github.com/remotion-dev/remotion/blob/main/LICENSE.md)), several popular model weights are for non-commercial use only, Artlist forbids automated downloads ([Artlist license](https://help-center.artlist.io/privacy-terms/artlist-license)), and some Pixabay tracks are registered with Content ID ([Pixabay FAQ](https://pixabay.com/service/faq/)). The skills default to generated music and effects, with no license to track. 06 ### Synthetic media Generated B-roll, avatars and cloned voices stay off. YouTube began labeling videos with significant photorealistic AI-generated content automatically in May 2026 ([The Next Web](https://thenextweb.com/news/youtube-will-now-automatically-label-ai-generated-videos-whether-creators-disclose-them-or-not)), TikTok requires clear labeling when AI or editing realistically depicts people or scenes ([TikTok Community Guidelines, via ConductAtlas](https://conductatlas.com/platform/tiktok/tiktok-community-guidelines/ai-generated-content-labeling-requirement/)), and the EU AI Act's transparency rules became enforceable on August 2, 2026 ([Faegre Drinker](https://www.faegredrinker.com/zh/insights/publications/2026/7/eu-ai-act-commission-confirms-transparency-code-of-practice-as-adequate-and-publishes-final-version-of-its-guidelines-on-transparency-obligations)). Run it yourself ## A test on one recording, then the level that fits each post. Everything below runs from this page or from the skills on GitHub. You keep every folder, every sign-in and every post. 1. **Give your assistant the method.** Use the Ask your AI bar at the top of this page, or install the skills: `npx skills add lkorczowski/content-strategist-AI-cookbook --skill statur-video-editor-pro`, and the same command for [statur-motion-design](https://github.com/lkorczowski/content-strategist-AI-cookbook/tree/main/skills/statur-motion-design), [statur-sound-design](https://github.com/lkorczowski/content-strategist-AI-cookbook/tree/main/skills/statur-sound-design), [statur-color-grade](https://github.com/lkorczowski/content-strategist-AI-cookbook/tree/main/skills/statur-color-grade), [statur-broll-find](https://github.com/lkorczowski/content-strategist-AI-cookbook/tree/main/skills/statur-broll-find). 2. **Write your context file once.** Your colors, fonts and what your brand never does, the numbers you may show in public with their source, the names that never appear, and the pages and footage you own. The editor skill reads it first (kb/01\_BRAND.md or .agents/product-marketing.md). 3. **Install the free tools.** ffmpeg, Python with OpenCV and numpy, Playwright with Chromium, the brand fonts and two model files, all listed in the skill. The models download from GitHub, which most sandboxes can reach. 4. **Test on one recording.** Ask for a level 1 and a level 3 version of the same clip, watch both on your phone, and judge on that clip alone. 5. **Then set the level per post.** Level 1 for clipping batches, 3 or 4 for a launch, and the level written in your video log next to each post, so your account learns which one works for you. Nothing is posted for you. For your AI ## The same method, written for your assistant. Skills in the open Agent Skills format, maintained on GitHub so every update reaches you. Upload them in Claude, paste them in ChatGPT, or install them with one command. [statur-video-editor-pro](https://github.com/lkorczowski/content-strategist-AI-cookbook/tree/main/skills/statur-video-editor-pro) The production dial, the intake (level, reference video, assets, music), one insert per sentence, the render and the quality gate that can refuse a clip. [statur-motion-design](https://github.com/lkorczowski/content-strategist-AI-cookbook/tree/main/skills/statur-motion-design) Hook box, captions with emphasis, cards, counters, page captures, diagrams and end card as HTML templates rendered with a transparent background in your brand. [statur-sound-design](https://github.com/lkorczowski/content-strategist-AI-cookbook/tree/main/skills/statur-sound-design) Voice cleanup, one effect per visual event, a music bed under the voice and a master at -14 LUFS, with a generated library that needs no license. [statur-color-grade](https://github.com/lkorczowski/content-strategist-AI-cookbook/tree/main/skills/statur-color-grade) Phone HDR to a clean SDR plate, the grade as a readable LUT, and measures of skin and levels. [statur-broll-find](https://github.com/lkorczowski/content-strategist-AI-cookbook/tree/main/skills/statur-broll-find) What to show for each sentence, captures of your own pages (signed-in pages through a snapshot) and an asset log with rights. [statur-vork-3d](https://github.com/lkorczowski/content-strategist-AI-cookbook/tree/main/skills/statur-vork-3d) Short 3D explainers and 3D inserts from presets written as code, rendered with vork on a GPU machine or with a CPU engine anywhere. **Rules every skill follows.** Drafts only: the AI never sends or publishes for you. No passwords: you sign in yourself. No promised views. What to expect The dial, the receipts and the gate are yours to keep. Views are not guaranteed: a better edit gives a good idea a fairer chance in the feed, and the idea still decides. Level 4 costs about 35 to 45 minutes of compute per 90-second clip on a small machine, so keep it for the posts that deserve it. The agent does the hours of editing. The ideas, the level, the final call and the post button stay with you. The story behind it I built the clipping chain first, for my own lives: [STATUR Clipping, the handbook](/clipping/). This one adds the production dial on top of the same cut, tested on a selfie of mine before I wrote a line of it. The research behind it also asked what happens to the job itself, [will AI replace video editors](https://louis-korczowski.org/ideas), and my answer is an essay by Dr. Louis Korczowski on louis-korczowski.org, coming out after this handbook. The format system behind the videos is on the [method page](/method/). Questions ## Editing with an AI agent, answered. What is an AI video editor? An AI agent that edits your own footage from written instructions: it cuts on the transcript, adds captions, graphics, captures of your pages and sound, renders the file and checks it. In this handbook it runs open skills with free tools, at the production level you choose. It prepares the edit; you review it and you post it. Can AI add B-roll, motion graphics and sound design to a talking-head video? Yes. The agent plans one insert per sentence that needs proof (a capture of your page, a list from your data, a stat with its source, B-roll from your own footage), renders the graphics as HTML templates frame by frame in your brand, and places one sound effect on each graphic entrance over a music bed. On my test it edited a 97-second iPhone selfie at level 4, with 124 sound cues and a master at -14 LUFS. Do I need a GPU or paid software? No. The default tools run on a CPU and cost nothing: ffmpeg, a headless browser for the graphics, OpenCV and numpy, and Robust Video Matting for the person matte. On 2 cloud vCPUs, the level 4 edit of 91 seconds took about 35 to 45 minutes of compute. Paid tools fit as options in their slot, once their license is checked. Which production level should I pick? Level 1 for clipping batches, lives and opinion posts. Level 2 for a single video with one message. Level 3 when you make claims a viewer can check. Level 4 for launches, flagship explainers and brand stories. With no data on your account, post a level 1 and a level 3 version of the same clip and compare. Do I have to label videos edited with AI? The rules target realistic depictions. TikTok's Community Guidelines require clear labeling when AI or editing is used to realistically depict people or scenes, YouTube began labeling videos with significant photorealistic AI-generated content in May 2026, and the EU AI Act's transparency rules became enforceable on August 2, 2026, with a clear label on deepfakes. If an edit makes a person or a place look real when it is not, label it. The skills keep synthetic media off unless you switch it on. Version 1.1 · last updated 2026-11-08: Version 1.1: a sixth companion skill, statur-vork-3d, for short 3D inserts and explainers written as code, and generated 3D inserts in statur-video-editor-pro 1.1.0. First edition: a four-level production dial on top of the clipping cut, nine recipes, a prompt per recipe and five companion skills in the public cookbook. Written from a real level 4 run on a 97-second iPhone selfie and a research pass on what viewers reward, with every number traced to its source. Changes are logged in the [changelog](https://github.com/lkorczowski/content-strategist-AI-cookbook/blob/main/CHANGELOG.md). Author: Dr. Louis Korczowski.