Method Guide · FreeVersion 1.0 · Miss AI
Miss AI · The Method
How that video
edited itself
The pipeline that cut the video you just watched. The numbers that control how it feels. The parts that don't work yet. The parts I'm keeping.
2:29 → 1:24 — 65 seconds of nothing, gone, without watching a timeline once.
Keira Nesdale · Miss AI · @RealMissAI
Free Guide
7 Steps
The exact numbers
Free tools
↓ Download PDF
Read first
Fair warning.
This isn't a product you can download. It's a method, built out of four free tools plus a few hundred lines of glue.
I'm giving you the method and the numbers, because that's the part that took me the time. Rebuilding the glue once you know the shape is a weekend, not a month.
If it turns out you'd rather not spend the weekend, there's a way to skip it at the bottom.
The Idea
The idea the whole
thing rests on.
The unit is the line, not the video.
That sounds like nothing. It's the entire trick.
If you record yourself rambling for three minutes and then ask software to find the good bits, you're asking it to guess. Guessing is why every auto-editor you've tried gave you clips that start halfway through a sentence.
If you write the lines first, read them to camera, fluff them, say them again better — then the software isn't guessing. It's matching what you said to the line you meant to say. That's alignment, and alignment is reliable.
Same footage. Completely different problem to solve.
How to Record
How to record so the
software can do its job.
This part is free and it's most of the win. Get the recording right and the rest is arithmetic.
One continuous recording. Don't stop and start. The whole point is that it finds every attempt at every line and keeps the best one, so messing up costs you nothing.
Fluff it on purpose. Say a line, hate it, say it again. Say it four times. More attempts means a better winner, and you're no longer performing, you're just talking until one comes out right.
Leave about a second of silence between takes. That gap is what lets the cutter find clean edges. It's the single thing most likely to make your first attempt go badly if you skip it.
Don't be word perfect. The matcher expects paraphrase, dropped filler and small reorderings. Say it how it comes out.
Talking between lines is fine. "Sorry, again" is fine. Junk at the top and the tail is fine. It all gets cut.
Skipping a line is fine. It comes back flagged as missing rather than quietly vanishing — which is a rule worth holding everywhere in a pipeline like this.
The video you watched was one pass in my kitchen with the phone propped up. Full frame, face fills it, no split screen.
The Method
The seven steps.
01
Write the script as numbered lines.
Not paragraphs. Lines, each tagged with what it's doing: hook, what-it-is, how, result, value, cta.
The tag does two jobs. It stops you recording something shapeless, because you can see when you've written four explanations and no payoff. And later it tells the graphics layer which asset is even eligible at that moment, so a callout can't turn up on your hook.
02
Read it to camera. Fluff it as much as you like.
Everything in the section above. Don't stop recording, don't do anything clever.
03
Transcribe with word-level timings.
WhisperX, free, runs on your machine. You need per-word start and end times, not sentence-level subtitles. Everything downstream keys off word edges, and a cut placed on a sentence boundary is exactly how you get clipped words.
First run pulls down a few gigabytes of models. Budget for that once and never think about it again.
04
Find every attempt at every line.
Score each stretch of transcript against each script line. The scoring has to tolerate you rewording things, because you will. One line in that video was scripted as "a skill doesn't, it's still there tomorrow" and I said "a skill will be inside your AI every single time you use it". Three words in common. Obviously the right take.
Watch for the stupid stuff: a transcriber writes "fifteen" one time and "15" the next for the exact same word. Numbers have to be flattened to one form on both sides before anything is compared.
05
Pick the winner in script order.
This is the bit people get wrong. Don't pick each line's best match on its own. Assign them in order, maximising confidence across the whole script.
Order is a stronger signal than word overlap, and it's what rescues a line you rewrote completely, because there's only one place it could sit. It also means overlaps sort themselves out: two lines can't claim the same seconds if the sequence has to hold.
The cost of that assumption, stated plainly: if a recording wanders badly out of script order, this is the thing that breaks.
06
Cut on word edges, then kill the silence.
Three numbers, measured off a working pipeline:
| Value | Name | Why |
|---|
| 0.04s | Lead | Padding before the first word, so it's never clipped. |
| 0.06s | Tail | More than the lead, because opening in dead air reads as a slow cut and ending a fraction late reads as breathing room. |
| 0.22s | Gap | The longest silence allowed to survive inside a take. This is the tightness dial. It's why the finished video runs at about 250 words a minute instead of 180. |
One consequence worth knowing: a single line often comes out as several separate pieces, because any silence longer than the gap gets removed by splitting the clip in two. Your line count and your clip count won't match, and that's correct.
On the video you watched: 2 minutes 29 down to 1 minute 24. Sixty-five seconds of nothing, gone, without watching a timeline once.
07
Stitch, then render the graphics on top.
ffmpeg for the stitch. Use trim in a filter graph, not -ss with stream copy: copy snaps to the nearest keyframe, which on a phone recording can be seconds away from where you meant to cut.
Then Remotion for everything on top. It's React, so a caption is a div and an animation is a CSS transform. If you can build a web page you can build these graphics — which is the whole reason I stopped looking for a video editor.
Anchor every graphic to a line number, never to a timestamp. "Callout on line 6", not "callout at 24.3 seconds". Change your mind about which take wins, re-run, and every graphic still lands where it should.
The Human In The Loop
The one place a
human still decides.
Every run spits out a page listing each line with the take it chose and every alternate it rejected, all of them playable. You skim it, you play the two you're unsure about, you override anything you'd rather have.
It takes about a minute and it's the only gate. Build one before you build anything fancy, because it's what turns "the software cut my video" from a thing you have to trust into a thing you can check.
It's also where the failures surface. A line that couldn't be matched shows up flagged and waiting, never as a silent wrong cut. That distinction is worth more than any amount of accuracy.
The Cost
What it actually cost.
| Tool | What for | Price |
|---|
| WhisperX | Word timings | Free, runs locally |
| ffmpeg | The stitch | Free |
| Remotion | Captions and graphics | Free for individuals |
| Claude Code | Wrote the glue, and does the directing | Paid plan |
The glue is a few hundred lines of Python doing the align, choose and cut, and some React for the graphics. Claude wrote nearly all of it, which is the joke inside the joke.
What I'm Keeping
What I'm not publishing,
and why.
You've got the shape of the whole thing. Three pieces of it I'm keeping, and I'd rather name them than have you assume I forgot.
The scoring
How a stretch of transcript gets matched to the line it belongs to, and the thresholds that decide when a match is good enough to win. That's the piece that took the longest and broke twice on real footage before it held.
The ordering method
Step 5 tells you to assign in order across the whole script. Doing that well, rather than greedily, is its own problem.
The graphics layer
The vocabulary of beats and the style library the renderer draws from. This is the part that makes it look like mine.
The reason
I'm turning this into a product, and a format stops working the moment fifty people run it identically. You can build all three yourself from what's here. I'd just rather you build your own version than paste mine.
Honest Limitations
What doesn't work yet.
I'd rather tell you this than have you find out.
×No finger tracking. "Follow my finger" in that video is a punch-in and a slow pan I placed by hand. It looks like tracking. It isn't.
×Short lines are fragile. A three-word line like "Zoom in here" can get swallowed by a longer overlapping match. It shows up as a line needing a manual pick, never as a silent wrong cut, but it does need a human eye.
×Nothing writes the script. Step 1 is still entirely me. That's the next thing I'm building.
×B-roll is manual. The software puts an asset on screen. Choosing a good one is still a person with taste.
Save Yourself an Afternoon
Two things that will
cost you an afternoon.
Both of these took me longer than they should have, and neither is written down anywhere obvious.
Give the render far more time than feels sane. Font loading runs on the render's overall timeout, so a long video with a couple of typefaces quietly runs out of road. The error it prints never mentions fonts. A short test render passing proves nothing about a long one.
A cancelled render poisons the next one. Stop a render partway and it leaves a half-finished cache behind, so the very next run fails with the same error for a completely unrelated reason. If a render breaks right after you killed one, just run it again before you believe a word of the message.
If you'd rather not build it
I'll set it up for you.
On your machine, with your look on the graphics, running on your own footage. I'll show you how to drive it so you're not calling me every time you make a video.
Book a call with Keira →
If you don't need me, I'll say so on the call and you'll still leave with a better answer than you came with.
If You Only Take One Thing
Script it in lines.
Everything else here is downstream of that decision, and it's the one that costs you nothing to try tomorrow.
Built in public. If you rebuild this, I'd genuinely like to see it.