toreel · How it works
The pipeline, timed
How a night of Reels actually gets made
Every number below comes from one real run: the subjects typed at night, fifteen finished reels scheduled by morning. Nothing was approved in between — and fifteen was that night's choice, not a fixed size.
Most tools that make short video with AI make one video. You write the prompt, you wait, you get a clip, and the next one starts from nothing. toreel takes the other unit of work: the night's queue. One reel a day is a new subject. Several in a day share one, so you can keep the angle that takes off.
The whole thing runs unattended. There is no step where it stops and asks you to approve something, because a pipeline that waits for a person is a pipeline that runs at the speed of that person being awake.
- Research 5 5 findings · 17 sources
- Scripts 15 one angle each
- Avatars 15 filmed, gaze corrected
- Montage 15 your footage, cut to the words
- Covers 15 typeset to fit
- Queue 15 spread across the day
A run, at about a thousand times its real speed. The real one takes two hours and happens while you are asleep.
Where the two hours go
One clock, from one finished run on 13 September. The two long bands are the two stages that are somebody else's GPU queue, and montage runs inside the covers window rather than after it.
Montage sits inside the covers window — the two run together, which is why seven stages totalling 157 minutes of work finish in 132 minutes of clock. The five-minute gap before covers is the batch handing over.
A night you asked for fifteen
The run measured here is a day of testing: one subject, fifteen angles, posted across one day. That is what five-a-day is for. A night of one reel is the same six stages on a new subject — not the same subject stretched until morning.
The six stages
Find out before writing
The subject is researched first — not summarised, researched. What comes back is a set of findings, each with its claims marked as confirmed fact, supported analysis or interpretation, and each with its sources.
measured: 2 min · 5 findings · 20 claims · 17 sourcesA script per reel, not a rewrite
Each reel gets its own script. On a one-a-day night that script is a new subject. On a several-a-day night the scripts are different arguments on the same subject — not one argument said five ways.
measured: 3 min · 15 scripts · ~600 s of speechFilmed without you in the room
Every line is delivered by your avatar, then passed through a second model that corrects the gaze to camera — the difference between someone reading and someone talking to you. Two engine steps, one idea.
measured: 51 min filming · 13 min gaze correctionYour visual source, timed to the words
The speech is transcribed, the words are timed, and your chosen visuals are cut against them. That source can be your own clips, licensed stock, generated stills or motion graphics. Subtitles are burned in at the size that survives a phone.
measured: 30 min · ~12 cuts per reel · 7 clips analysed into 53 usable segmentsA cover that still reads at thumbnail size
A frame from the montage becomes the reference, a background is generated around it, and the headline is typeset to fit — measured against the box it has to live in rather than hoped into place.
measured: 54 min · 15 covers · 0 refused for overflowSpread across a day, not dumped
Fifteen reels posted at once is a wall nobody scrolls. They are spread across the posting day on an hour pattern, each with its own minute, and handed to the publisher so the queue survives your laptop closing.
measured: 4 min · 15 jobs · first at 01:54, last at 23:05What fails, and what happens then
Reels fail. An avatar comes back broken, a render runs out of something, a model refuses a prompt. The rule is that one bad reel does not cost the batch: a reel is retried, and if it keeps failing it is dropped and the other fourteen carry on. You wake up to fourteen, not to nothing.
The same rule applies upward. A stage is only finished when something actually came out of it — a stage where every reel failed is a failed stage, not a finished one, however many things it tried.
Why it runs at night
Look at the timings again: filming is 51 minutes, covers 54, montage 30. Almost none of that is your computer thinking — it is queueing behind other people's GPUs. There is no version of this that returns fifteen finished reels in five minutes, so the honest design is to admit it takes hours and put those hours where you are not waiting through them.
And then the part most tools stop before
A batch that posts and is never looked at again is content, not a system. When several go out the same day, they go out as a trial — same subject, different angles, keep the one that takes off. One a day is publishing. What comes back still decides what happens next.
Several in a day go out as a trial
Same subject, different angles, spread across the day rather than dumped. Nothing about them is precious — the day exists to find out which angle your audience actually wanted. One a day skips this and just publishes.
The numbers come back on their own
Views, reach, saves, shares, how long people watched before leaving. Pulled from the account automatically and attached to the reel that earned them, so a result is never separated from the thing that produced it.
What landed gets a second life
The winners are the seed of the next batch: the angle that worked becomes the next subject, the hour that worked becomes the hour, the cover that worked becomes the pattern. The rest is not failure, it is the measurement that told you.
Today the first two stations are built: the batch posts on a schedule, and the numbers come back by themselves — 268 reels of the author's own account are already being read this way. Promotion is what is being built now: taking the reel that landed and making it the seed of the next subject rather than a screenshot in a folder.
What it did to one account
The only account that has run on toreel so far is the author's. One account is not proof of a market, so here is the whole of it rather than the flattering part.
Reels published per month on one account. Blue is by hand; the lighter bars are months where toreel wrote, filmed and cut most of them. September is a partial month.
The month toreel started, output went from 58 reels to 126 — and views per reel did not go up with it. Compared like with like, only reels published after the pipeline's first one on 10 August:
| Since 10 Aug | Reels | Median views | Top 10% | Avg watch |
|---|---|---|---|---|
| Made by toreel | 127 | 1,703 | 3,709 | 12.6s |
| Made by hand | 19 | 1,893 | 4,678 | 13.1s |
The machine holds the median and multiplies the volume. It does not beat a person on a single reel — the median is ten per cent lower and the watch time is within half a second. What it changes is how many there are: nineteen by hand in a month against a hundred and twenty-seven.
The honest caveat, because it matters: every reel that has gone viral on that account so far was made by hand. The machine keeps the middle of the distribution. It has not yet produced the top of it.
Why the median was never the point
Sort those 268 reels by how they did and the shape is not a bell. It is a cliff — and it is the reason to care about how many reels you make rather than how good any one of them is.
Share of all views, across 268 published reels on one account. Median reel: 2,102 views. Best reel: 839,739. The three shares add to 100. The distribution is the product argument, not a detail of it.
Twenty-seven reels out of 268 carry four fifths of everything. The other 241 were not bad; most were fine. Which one would land was not knowable before posting it, and nobody — no tool, no strategy, no hook formula — knows in advance either.
That is the whole argument. If the outcome is decided by something you cannot predict, the only lever you actually hold is how many times you enter. Fifteen a night against two a week is not a productivity improvement; it is a different number of chances at the only part that matters.
It also explains the one honest weakness above. The machine matching the median is not the interesting question — the median reel earns nothing either way. The question is whether it eventually produces one of the twenty-seven, and the only way to find that out is to keep entering.
What the data refused to say
The same 268 reels were checked for a best hour to post. Across the nineteen hours holding five reels or more, the medians run from 325 to 3,724 — and eighteen of those nineteen sit inside one band, in no order at all.
So toreel does not claim one. A batch is spread across the posting day because fifteen posts in one hour reads as a dump, not because any hour was found to be magic. Anyone selling you a posting-time secret from a sample this size is reading noise.
Eighteen of the nineteen sit inside one band, in no order. The nineteenth is 20:00, which is seven reels — a sample that small moves further than any hour of the day could.
Type a subject, sleep on it
toreel is not open yet. Leave an address and you will hear the day it is.