Your sharpest answers come out live, when someone asks the question you never planned for. Ten clips cut by hand from a one-hour webinar take one to two and a half working days, so most of those answers stay buried. This handbook runs the chain with an AI agent and free tools on your own computer, from a live, a webinar, a long YouTube video of yours or a Google Drive folder: transcribe, check the sync, plan the clips, edit, review from your phone and file the masters. Every recipe ends with one thing to do today and a prompt for your assistant.
Record your live on your own machine: TikTok and Facebook delete live replays after about 30 days, and the TikTok and YouTube Studio downloads stop at 720p. A webinar file, a long YouTube video of yours or a whole Google Drive folder works too. Transcribe on your computer with word timestamps, in French or English. Check that the transcript lines up with the video. Read the whole transcript with your AI and score every clip candidate on six criteria. Write each edit as a short spec, render it where the footage lives, at any length and resolution, and let a re-transcription catch the bad cuts. Review on your phone with notes at the exact frame. File the masters with one naming rule. The agent does the hours. You keep the ideas, the final call and the post button.
A live forces you to think out loud, in front of the people who ask the real questions. The value is there. The hours to cut it are missing.
Goldcast counted 102 live attendees and 14 on-demand views per B2B webinar on average, across 26,190 webinars in 2025, and about a third of the webinars offered on demand got no on-demand view (calc.) (Goldcast, 2026 report, vendor data).
Wistia published its own numbers: the clips of its 2025 webinar passed 800,000 views, and the live itself made under 0.1% of the views (Wistia, August 2025, vendor case).
Fiverr counts one to two hours of editing per finished minute (Fiverr guide), and an editor produced five to ten polished clips a day before AI tools (ClipSpeed, vendor): ten one-minute clips take 10 to 20 hours (calc.). Most social video editors on Fiverr deliver in about 11 days (Fiverr).
TikTok and Facebook keep replays about 30 days, and the TikTok and YouTube Studio downloads stop at 720p (recipes 01 and 02). Shoots recorded on purpose pile up the same way: the footage exists, and nobody has the hours to cut it.
Clients, product, hiring and investors all land on the same week. The clipping backlog competes with everything else, so the best answers of the live stay buried.
Nine recipes. Each one fits on a phone screen, ends with one thing to do today, and comes with a prompt for your AI assistant.
| When | What | Time |
|---|---|---|
| During the live | Record locally | no extra work |
| Right after | Transcribe and check the sync | about 15 min of machine time per 100 min of live |
| Same day | Plan the clips with your AI | one session |
| Per clip | Spec, render, quality check | about 1 min of rendering, under 1 min of checks |
| Per batch | Review on your phone | a few minutes per clip |
Every number below was measured on that run. The same live gave 10 clips.
| Step | What happened |
|---|---|
| Live | 1:40:01 on TikTok LIVE, in French |
| Replay | 960x1920 from my own account, 2.7 GB. TikTok's own LIVE Center download offered 720p after about 1 h 30 of preparation |
| Transcription | About 17,000 words with timestamps, in about 15 minutes for 100 minutes of audio on 2 cloud vCPUs |
| Sync check | audio 1.71 s late on the video, drifting 0.19 s over the live: measured and corrected |
| Clip plan | 17 candidates, 16 of one minute or more, 14 scored |
| Clips | 10 clips of one to two minutes, each with a one-sentence hook, captions and a moving watermark, two of them with chapter labels |
| First clip | 1:28, 26 pieces, rendered in 72 seconds on a 4-core virtual machine of my Mac, 1080x1920 at -14 LUFS |
| Quality check | a 97.7% word match with the plan (361 of 371 planned words heard) |
| Review | one change (the on-screen hook), sent from my phone about 4 minutes after the link; version 2 rendered in 44 seconds |
The platform replay is a backup. Your master is the file you record yourself, at the resolution you stream.
YouTube recommends keeping a local backup of a live and warns that a stream longer than 12 hours may not be archived at all (YouTube Help). Platform downloads also come back smaller than the stream: YouTube Studio offers MP4 files at 720p or 360p (YouTube Help).
Copy this prompt into your assistant:
I stream on [platform] with [tool] from a [computer]. Tell me how to record every live locally at the resolution I stream, where the file lands, how much disk space a two-hour live needs, and how to test it in one minute.
Skill: statur-clipping · install: npx skills add lkorczowski/content-strategist-AI-cookbook --skill statur-clipping
No local file? Get the platform's own download the same week, or point the agent to the file where it already lives: a long YouTube video of yours, a Google Drive file or a whole Drive folder.
Facebook keeps live videos for 30 days since February 2025 (Meta, February 2025). TikTok replays stay visible to the creator for about 30 days according to a 2026 agency guide (TSL Agency, October 2026). YouTube's terms allow downloads only through YouTube's own features or with prior written permission (YouTube Terms of Service): a third-party download tool sits outside them, even for your own videos.
Copy this prompt into your assistant:
List my lives and long videos of the last 30 days on [platforms] and in my Google Drive folder [folder]. For each one, tell me how to get the file with the platform's own tools or from Drive, at what quality, and until when it stays available. Then add them to my clipping queue, one project per video.
Skill: statur-clipping · install: npx skills add lkorczowski/content-strategist-AI-cookbook --skill statur-clipping
Word timestamps are the backbone of every cut. Get them locally, with a model whose French and English quality has been measured.
In the Open ASR Leaderboard paper (version 4, March 2026), Parakeet TDT 0.6B v3 scores a 5.42% word error rate in French, against 6.36% for Whisper large-v3 (arXiv 2510.06961). It runs on a CPU, without a GPU or an API key: on my live it transcribed 100 minutes in about 15 minutes on 2 cloud vCPUs.
Copy this prompt into your assistant:
Help me install a local transcription with word timestamps on my [Mac, Windows or Linux] computer: sherpa-onnx with NVIDIA Parakeet TDT 0.6B v3 and Silero VAD, from the sherpa-onnx GitHub releases. Then transcribe [file] and save words.json and a readable transcript in [folder].
Skill: statur-clipping · install: npx skills add lkorczowski/content-strategist-AI-cookbook --skill statur-clipping
One offset moves every cut. Measure it before the first edit.
On my live, the TikTok replay started its audio 1.71 seconds after the video and drifted 0.19 seconds over 100 minutes. My first render had every cut early. The check takes a few seconds, and every clip rendered with its map landed on time.
Copy this prompt into your assistant:
Compare the audio of [video file] with my transcription audio [audio file] at 20 points across the recording. Report the offset at each point in milliseconds and tell me whether cuts made from the transcript will land on time. If they will not, write a sync map I can pass to the renderer.
Skill: statur-clipping · install: npx skills add lkorczowski/content-strategist-AI-cookbook --skill statur-clipping
Read the whole transcript before you choose. A strong answer can sit in the last twenty minutes, and a question asked at minute 12 can get its answer at minute 47.
Live answers come in pieces. A viewer asks, you answer, and you come back to it twenty minutes later. A clip that stitches the same question across the live can beat any single passage. My 1:40 live gave 17 candidates.
Copy this prompt into your assistant:
Read my company context file [file] first. Here is the transcript of my live with timestamps: [file]. Map the topics with their time ranges, list every clip candidate of 60 seconds or more (stitch scattered answers, overlaps allowed), apply the hard gates, score each one on hook, clarity, value, payoff, delivery and fit, tag it TOFU, MOFU or BOFU, and rank them.
Skill: statur-clip-plan · install: npx skills add lkorczowski/content-strategist-AI-cookbook --skill statur-clip-plan
The edit is a short JSON spec: passages in any order, cuts, caption fixes, hook and framing. The renderer does the rest, at any length and resolution.
On my live, a 1:28 clip of 26 pieces rendered in 72 seconds on a 4-core virtual machine of my Mac, at 1080x1920 and -14 LUFS. The one change I asked for in review, a new hook, rendered again in 44 seconds.
Copy this prompt into your assistant:
Write the edit spec of clip [id] from my words file [file]: passages in play order, cuts, caption fixes, a one-sentence hook of ten words or fewer in my voice, chapter labels and framing. Then render it with clip_render.py at 1080x1920 and tell me its duration and number of cuts.
Skill: statur-clip-edit · install: npx skills add lkorczowski/content-strategist-AI-cookbook --skill statur-clip-edit
Transcribe the finished clip again and compare it with the words you meant to keep.
While I built the chain, the check caught two problems before I saw anything: the 1.71-second audio offset of the replay, and a fade that muted most of a test render. The clip I received matched the plan at 97.7% (361 of 371 planned words heard), at -14 LUFS.
Copy this prompt into your assistant:
Run the quality check on [master file] against [planned words file]: transcribe it again, align it with the planned words, list every difference near a cut, and report loudness, true peak, duration, resolution and frame rate. Tell me if it is ready for review.
Skill: statur-clip-edit · install: npx skills add lkorczowski/content-strategist-AI-cookbook --skill statur-clip-edit
The founder's time is the scarcest input. Give one link, a player, a note at the exact frame and three verdicts.
I reviewed my first clip on my phone and sent one change, a new hook, about four minutes after receiving the link. Teams that already review in Frame.io can keep it: it offers comment webhooks and an MCP server in public beta (Frame.io webhooks, Frame.io MCP server).
Copy this prompt into your assistant:
Build a private review page for these clips [files]: a 9:16 player, timecoded notes with tags, a verdict per clip (ready, changes, drop) and the brief of each clip. Then read the notes and turn each one into a change in the clip spec.
Skill: statur-clip-review · install: npx skills add lkorczowski/content-strategist-AI-cookbook --skill statur-clip-review
One folder per live with three subfolders, and one name per master that says the live, the clip and the version.
Ten clips with two versions each make twenty files after a single live. A name that carries the version, and an archives folder for everything older, keep the approved one easy to find: whatever sits at the root is ready.
Copy this prompt into your assistant:
Create the project folder for my live of [date] with RUSHS, TRAVAIL and EXPORTS, move the source into RUSHS, and give me the naming rule for every master, caption file and review copy.
Skill: statur-clipping · install: npx skills add lkorczowski/content-strategist-AI-cookbook --skill statur-clipping
Assumption: one live or webinar of 60 to 120 minutes a week, so 4.33 lives a month (calc.), and about ten clips per live. Prices read on October 24, 2026. (calc.) marks my arithmetic on sourced numbers.
| SaaS clipping tool | Editor or agency | Local agent pipeline | |
|---|---|---|---|
| Per month | $10 to $68 billed annually, $24 to $78 billed monthly (calc.), with Reap, StreamYard, Riverside or Restream plans | $1,733 to $3,467 for a mid-level Upwork freelancer (calc.); $2,500 to $10,000 a month at Clipping, the agency behind MrBeast's clips | $0 in licenses for the open-source tools, plus your AI assistant plan and your own computer |
| Per live | about $2.30 to $15.70 billed annually (calc.) | $400 to $800 with a mid-level Upwork freelancer (calc.) | close to zero at the margin, outside the AI assistant |
| Time per live | about 5 to 10 minutes per hour of video at Reap (vendor figure) | 10 to 20 hours of editing (calc.) | about 15 min of transcription per 100 min, about 1 min of rendering per clip (measured) |
| Delivery | minutes, plus your review | about 11 days for most social video editors on Fiverr; 3 days at a done-for-you clipping service | the same day, plus your review (measured) |
| French quality | listed by every tool; I found no independent test of their French | depends on the editor | model benchmarked at a 5.42% word error rate in French (Parakeet TDT 0.6B v3; the pipeline runs its int8 build) |
| Rights on your footage | some terms grant the tool a perpetual, sublicensable license on uploads | assign the rights in the contract | the video files stay on your machine; the transcript goes to your AI assistant and the light review copies to your private review page |
| Left to you | review, approval, guests' consent | review, approval, guests' consent | review, approval, guests' consent, and the selection brief |
Sources: Reap, StreamYard, Riverside, Restream and OpusClip pricing pages; Upwork editor rates (median of $40 an hour at mid level, applied to 10 to 20 hours); Fiverr guide and Fiverr delivery data; Clipwing (3-day delivery); Complex, November 2025 on the clipping agency; OpusClip terms; Open ASR Leaderboard. Measured: my live of October 24, 2026.
The transcript drops most "euh" and false starts, so a few survive the cuts. Budget a human pass on the clips that matter most.
The renderer crops one fixed frame. A guest in a split screen needs a crop per speaker.
No public source measures how long a founder takes to review clips. Mine took about four minutes for the first one. Measure yours, because it decides how many clips a week you can ship.
Get the written consent of every guest before you clip and republish them. A viewer's name or handle stays off screen unless there is a reason and their consent.
YouTube's monetization policy asks for your original creation and excludes reused content without a meaningful difference, even with the creator's permission (YouTube). Meta reduces the distribution of reposted content, even when the reposter adds a watermark (TechCrunch, July 2025), and Instagram makes Reels that carry another app's logo or watermark less discoverable (CNN, February 2021). The statur.co/clipping watermark is a small text credit on your own clip. No platform says how it treats a small text credit like this one: test a few clips with and without it. Use your own recording or the platform's own download.
A platform download can carry weaker audio and a smaller picture than your stream (720p on TikTok and YouTube Studio), and its audio can start late. Record locally for the lives that matter, and run the sync check on every file.
My rule of thumb after running this chain on my own lives: a skill without a clean knowledge base of your company delivers about 10% of its potential. The skills cut, caption and render. Picking the right moments needs your company: who buys, what you sell, which numbers you can say in public, which names never appear, which formats already work for you. Without that file, the plan picks entertaining moments. With it, the plan picks the moments that sell your ideas. Write the context file once (kb/01_BRAND.md or .agents/product-marketing.md), keep it current, and every skill reads it first.
Everything below runs from this page or from the skills on GitHub. You stay in control of every folder, every sign-in and every post.
npx skills add lkorczowski/content-strategist-AI-cookbook --skill statur-clipping (same command for statur-clip-plan, statur-clip-edit and statur-clip-review).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.
The whole chain for one recording or a batch: permissions, sources (lives, YouTube, Google Drive, local), transcription, sync check, queue, watermark, delivery and the weekly routine.
Read the full transcript, map the topics and score every clip candidate with the hard gates and six scores.
Turn a clip spec into a master at any length and resolution: cuts, captions, hook, loudness and the quality check.
Run the review loop on a private page with timecoded notes and a verdict per clip, then render the next version.
I started running lives in 2024 with Tonton Giovanni, a 25-seat pizzeria owner who grew from 2k to 1.6M followers, and the lives were part of that growth plan. Then I built this machine for my own lives. Read the story by Dr. Louis Korczowski on louis-korczowski.org. The format system behind the clips is explained on the method page.
Any recording you can get as a video file: TikTok LIVE, Instagram Live, YouTube Live, LinkedIn Live, Twitch, Zoom or a webinar platform, a long YouTube video of your own, or a Google Drive file or folder. A folder becomes a queue, one project per video. The chain starts from the file.
Yes. Parakeet TDT 0.6B v3 covers 25 European languages with punctuation and word timestamps, and its French word error rate is measured at 5.42% in the Open ASR Leaderboard paper. Captions come out in the language spoken.
The clip plan keeps clips between 60 seconds and 5 minutes. The renderer has no length or size limit and renders 4K when the source allows it. Only the review copies stay under 20 MB, so they open fast on a phone.
No GPU is needed. On my live, transcription ran on 2 cloud vCPUs and rendering on a 4-core virtual machine of my Mac. A recent Mac should be faster; I have not measured it.
The agent prepares drafts: masters, captions, post text, hashtags and pinned comment. You review, approve and post. It never posts, schedules or deletes anything for you.