Skip to content
Skip to content
How it works

How Studio works
and why you can check it.

Eight named stages run in a fixed order. Each one writes down what it did and why, and the times below are the times this server actually measured.

Every stage has a name, not a spinner Stopping a job charges nothing A job survives a server restart

The times below were measured on a 30-minute interview on this server.

Same order every time Eight steps. You can read each one while the video is made. Read the uploads, study your example, hear every word, cut retakes and pauses, decide every visual, draw and render, check the result, compare with your own edit. The pipeline walks that list and skips the entries that do not apply to your footage.
The check runs before the drawing No card will ever show a word you did not say. The word check runs inside the plan stage, before a single frame is drawn. A card carrying an unspoken word is dropped, and that sentence is planned again on your face instead — in the plan, not in the post-mortem.
A real edit, end to end

One 29.8-minute two-person call, edited on this server.

This is the job the stage times come from — a real recording, not a test clip. It went through the live Studio twice, its frames were looked at, and five faults in it became rules.

30 s to read the uploads — probe every file, measure the kind of footage, set each file's role.timed on this server
6 min to hear every word — a word-level transcript of the whole call. timed on this server
25 min to cut retakes and pauses — and, in a conversation, to keep the repeats that were meant.timed on this server
43 min to draw and render the video — the longest stage by a wide margin. timed on this server
Add the stages up and you get about 75 minutes
30 seconds, 6 minutes, 25 minutes, 30 seconds, 43 minutes, 5 minutes. Out came a 29.0-minute episode with 13 cards — about 2 % of the running time — and every gate that can block a video passed on the first finished version.
Where the checks run The gates are run against the live server, not in a lab. Each gate is a command with an expected output and the evidence it produced, run against the deployed Studio over its own URL. Ten test clips go through the live URL on every release and all ten must pass the same six blocking checks a customer's video passes.
On screen while it runs

Steps, the video, and the decisions, side by side.

While a video is being cut you see the stage it is on, the picture, and why each choice was made.

studio.itskaivalya.com/studio
Steps 4 of 6
✓Read the uploads30 s
✓Hear every word6 min
✓Cut retakes and pauses25 min
◐Decide every visualrunning
5Draw and render the video~43 min
6Check the result~5 min
The video about 34 min left
Ask for changeseach message makes a version
drawing chunk 2 of 4 · frame 41 of 96
11:04 / 29:02
What the editor decided
00:02:14 Two people in side-by-side call panels — treating this as a conversation The picture, after the black bars come off, is 3.6 : 1. The conversation cut and plan rules follow.
00:12:02 Dropped a card and planned its sentence back on the faces The word check found a word in the card that had not been spoken. The card was removed before anything was drawn. unspoken word — dropped
The eight stages

What happens to a 30-minute interview.

These are the steps you see on the video page. The times are from one real interview on this server.

Read the uploads Every file is opened and measured: its length, its shape, whether a face is in at least half of twelve sampled frames, whether the picture is as flat as a screen, how often the scene changes. Each file's role is then set for you — auto, talking, b-roll, image, logo or music. About 30 seconds. You see each file’s length and its role.
Study your example Runs only if you uploaded an example video. Its pace, caption style, cut rhythm, zoom strength, colour and sound are measured and applied to your footage. About 1 minute, and only if you uploaded an example.
Hear every word The speech is transcribed with word-level timings. A script you uploaded is used here: names and brands are spelled the way you wrote them, and only where the fix sounds like what was said. About 6 minutes. You can read the transcript.
Cut retakes and pauses A repeated take goes and the last complete one stays. In a conversation only stutter restarts go and the repeats are kept, because people repeat themselves on purpose. Sentences past 12 seconds are split at their natural breaks. About 25 minutes. You see what was kept and what was cut.
Decide every visual Every card, zoom, caption and pause is planned against the words that were really spoken. The word check runs here: a card holding a word nobody said is dropped and its sentence is planned again on your face, before anything is drawn. About 30 seconds. Each choice is written down with its reason.
Draw and render the video The longest stage by a wide margin: the video is drawn frame by frame and encoded. The drawing grows about 0.27 MB a frame, so on a long video the renderer starts itself again with fresh memory once it passes 1.5 GB — the finished chunks are kept. About 43 minutes. You see the percent done and the time left.
Check the result The finished file is opened and checked against the plan: it plays, it is the planned length, captions match the speech, no card covers a face, no card shows a word nobody said, the audio is not clipped or silent, the frame has no black edges. About 5 minutes. You see what passed and what needs a look.
Compare with your own edit Runs only if you uploaded your own cut of the same footage. Studio measures both edits and leaves the differences on the page; it does not change its own edit to match yours unless you ask. About 30 seconds, plus a little more for a long video. You see the differences.
Only the stages that apply are run
The pipeline asks which stages this upload needs and queues only those. A montage of clips, pictures and music has no transcription stage at all — nothing is spoken. A job with no example video has no style stage, and one with no reference video has no compare stage.
The four upload options

Four optional things you can put in with the footage.

Footage is the only thing a job requires. Everything below changes how the edit is made. Two of them must be videos; the other two are text.

The four upload options, what each one does and what it must be
OptionWhat it doesWhat it must be
Your own edit
reference
Studio measures your cut and its own and leaves the differences on the page. It keeps its own edit unless you ask it to follow yours — you are shown the gap rather than quietly overruled. A video. Pictures and sound files are not accepted here.
An example video
example
The example's pace, captions, cuts, zooms, colours and sound are measured and copied onto your footage. It copies the craft, never the content. A video, edited the way you want yours to feel.
A script
script
Names and brands are spelled the way you wrote them. A correction is made only where it sounds like what was actually said, so a script cannot put words in your mouth. Text, up to 200,000 characters.
A brand colour
accent
Replaces the amber accent used on cards, captions and the hook text, so the video matches the rest of your channel. Six hex digits, with or without the leading hash.
Measured, not imitated The example is measured with a ruler, not described in a sentence. Average shot length, words per caption, where the cuts fall, how hard the zoom punches, the colour grade and the music level all become numbers, and those numbers are applied to your footage. The measurement appears in your video's decisions in plain words.
The two brains

Two ways of deciding, switchable without a deploy.

The editor under the stages has two versions. They do not change what the stages are called or how long they take; they change how the visual plan is arrived at.

v1

The rules of the older Studio

The rulebook from version 0.6.3 and before: if this, then that, applied directly. It is guarded by ten regression clips — every release has to leave those ten edits unchanged.

“0 of 10 changed” on every release — the checks behind it

v2

Understand first, then show the structure

An understanding stage works out what the video is about before a visual is planned, and keeps a catalogue of the visuals it may use. The plan is made against that understanding, not sentence by sentence.

the understanding stage keeps a catalogue of the visuals it may use — the three sources, in order

The switch needs no deploy
The owner changes it on the live server and the next job started uses the other brain. There is no release, no restart and no lost queue — a job already running finishes on the brain it started with. Both brains share the same eight stages, the same cut rules, the same word check and the same checks on the finished file.
Nothing is lost

A dropped connection, a refresh, a restart — the job continues.

Long jobs meet real life: trains, hotel wifi, a server update at two in the morning. Each of those was treated as a bug to design out, and every fix is a rule with a test behind it.

What happens to a job or an upload when something goes wrong
What happensWhat Studio doesWhere it is proved
An upload is sent Files arrive in 4 MB chunks, up to five in parallel. A chunk that arrives ahead of the file's end waits and is joined the moment the bytes before it land, so an out-of-order piece is never a lost piece. One stream measured 0.31 MB/s; several measured 2.13 MB/s. UP1 · 0.31 MB/s became 2.13 MB/s
The connection drops, or you refresh The upload waits and continues by itself; nothing received is thrown away. The server kept every byte, so choosing the same files again continues from where it stopped. the upload continues by itself
The server restarts mid-job The job continues from its last saved stage or piece, up to five times. An update restart does not use up one of those attempts — only a crash does. DS2 · the decision log
A stage stalls for 45 minutes It is stopped and continued from the last piece it wrote, up to three times. Nothing already finished is done twice. the job's own stalled-stage record
A job runs and runs No job may run longer than 24 hours. If one ever does, it stops as failed and you get the full report rather than a page that never changes. the 24-hour limit, in the honest limits table
If something does go wrong

Two buttons, and one plain sentence about what happened.

Column 2 shows one plain sentence about what stopped the job, and the two things you can do about it. The technical traceback stays in the pipeline log, where it belongs.

ER1 · one plain sentence, and the traceback in the log

Try again from where it stopped Ask about an error

Stopping a job keeps everything done and charges nothing.

An honest clock

A real time, and it rounds late on purpose.

Studio measures how long each stage actually took on this server, fits it to the length of your footage, and adds up the stages still to come. That is the number next to your video.

# seconds = a + b × minutes of footage — fitted from this server's finished jobs talk probe (20, 1) transcribe (30, 12) cut (25, 50) plan (5, 1) render (60, 87) check (10, 10) montage probe (40, 110) transcribe (0, 0) cut (0, 0) plan (0, 0) render (100, 110) check (15, 5) # before the footage is read, its length is guessed from the upload size MB_PER_MIN = 60.0 # better a little late than early: the measured spread was −15 % to +40 % MARGIN = 1.15

A queued job adds the work of the jobs ahead of it, so a small reel is not told to wait for someone else's two-hour podcast. Once a running stage reports its own progress, its own pace beats the history. A stage past its usual time never shows zero minutes left: it keeps a minimum of its own estimate, or 20 seconds.

The 15 % The estimate is deliberately 15 % longer than the model says. Measured against finished jobs, this server's own predictions came out between 15 % early and 40 % late. When a number has to be wrong, the owner chose the direction that does not leave you refreshing a page: better to be told 75 minutes and get 68 than to be told 60 and still be waiting.
Why it exists at all The owner once waited an hour with “working…” and no idea when it would end. That sentence is written at the top of the file that now does the arithmetic. The honest clock was not a marketing idea — it was the fix for a page that said nothing while a long job ran.
After the edit

What you can change afterwards, in your own words.

Under the finished video there is one box. You write a sentence, Studio answers with exactly what it understood, and a new version is made. The earlier versions stay, so you can always go back. Write “always”, “in all my videos” or “from now on” and the wish becomes your account's default; write “for this project” and it stays with that project.

no captions fewer cards remove the card at 1:20 cut 2:10 to 2:30 faster make it vertical lower the music brighter

When Studio cannot work out what you meant it says so, and offers the real examples it does understand. When what you asked for is not possible it says that too — “Not possible yet”, with the part it could not do, rather than a silent no.

The two replies are “Got it: …” and “Not possible yet: …”. Each line of the learned record is dated and quotes the words that taught it, and what it learned is in the documentation.

Questions

The parts people ask about before their first long job.

Does every kind of video go through the same eight stages?
The list is the same, but only the stages that apply are queued. A montage has no transcription and no cut stage — nobody speaks. No example video, no style stage; no reference video, no compare stage.
How long will a 30-minute video take?
About 75 minutes on this server, measured: 30 seconds to read the files, 6 minutes to transcribe, 25 minutes to cut, 30 seconds to plan, 43 minutes to render, 5 minutes to check — fitted to your footage and multiplied by 1.15.
Can I watch the edit while it is still being made?
Yes. Column 2 shows each stage with its time and progress. Column 3 fills up with the decisions as they are taken — what kind of footage it saw, what it cut, and every card with its reason.
What happens if I close the tab?
The edit runs on the server, in its own process, so closing the tab does not stop it. The one thing to keep open is an upload in flight: the browser asks before closing the tab while bytes are still on their way.
What if the server restarts in the middle of my job?
The job continues from its last saved stage or piece, up to five times. An update restart does not count as one of those attempts — only a crash does. A stalled stage is continued from its last saved piece, and a job short of memory waits a minute and carries on.
Do I get charged if I stop it?
No. Stopping keeps everything already done and charges nothing, and minutes are only ever charged once, when a video finishes.

Watch your own footage go through it.

Your first 30 minutes are free — one real episode, edited end to end, with the steps and the decisions on screen the whole way.

Already have an account? Sign in.