How tondi picks which moments to clip
Checked against v1.1.5, 12 September 2026. The weights below are read from the code, not paraphrased.
tondi is not a language model, and it doesn't pretend to know what will go viral. It scores every stretch of whole sentences on a fixed set of signals — how loud the delivery is, whether the first words hook, how many numbers and superlatives it contains, how fast the speech is, how long it runs — and keeps the highest-scoring stretches that don't overlap. This page lists the signals and their weights.
What happens to your video, in order
- Transcribe. The audio is transcribed on your Mac with a timestamp on every word. Nothing is sent anywhere.
- Group into sentences. Words are grouped at full stops, question marks and exclamation marks. A pause longer than 1.2 seconds, or a run of 30 words, also ends a group — but those are marked as unfinished, because people pause mid-thought.
- List every candidate. A candidate is any run of consecutive sentences between the shortest and longest clip length you chose (defaults 18 and 55 seconds; the sliders run 10–30 and 30–90). It has to start right after a finished sentence and end on one, so a clip never opens or closes mid-thought.
- Score each candidate. The table below.
- Pick the best that don't overlap. Highest score first; anything within 1.5 seconds of a chosen clip is skipped. Before a clip is kept, a leading wrap-up sentence or a trailing "next up" sentence is trimmed off, as long as the clip stays above the minimum length.
- Pad the edges. 0.2 seconds before the first word and 0.35 after the last, but never into the next word, so captions don't show a word you can't hear.
The signals and their weights
Every candidate starts at zero. Points are added or removed as follows; the numbers are the constants in the code.
| Signal | Points | What it's looking for |
|---|---|---|
| Audio energy | −1.0 to +2.0 | How loud the stretch is compared with the rest of the recording. Excited delivery scores higher. |
| Hook opening | +1.2 | One of the first four words is a hook word — why, how, wait, imagine, secret, never, best, worst, number, finally… |
| Opens with a question | +0.5 | A question mark in the first sentence. |
| Hook words throughout | +0.25 each, up to +2.0 | The same list, anywhere in the stretch. |
| Superlatives | +0.3 each, up to +1.2 | Words ending in -est: biggest, fastest, cheapest. |
| Numbers | +0.15 each, up to +1.0 | Digits or number words. Rankings and lists clip well. |
| Punch | +0.25 each, up to +1.0 | Exclamation and question marks. |
| Speech rate | −0.5 to +1.0 | Words per second against the recording's average. Faster than usual scores higher. |
| Length | up to +1.0 | A bell curve centred on 32 seconds. A 16-second or 48-second stretch gets about a third of the bonus. |
| Starts a section | +1.0 | The first sentence opens like a new item: "number three", "next up", "moving on", "last but not least". |
| Straddles a section | −1.5 each, down to −3.0 | A section opener inside the stretch means it runs from one topic into the next. |
| Opens on a wrap-up | −1.0 each | One of the first two sentences is "and that's why…", "so yeah…", "at the end of the day…" — the tail of the previous topic. |
| Unrelated first sentence | −0.8 | The first sentence shares no content words with the rest — probably leftover from the previous point. |
What this is good at, and what it isn't
It works best on talk with structure: rankings, explanations, interviews, anything where a point starts, builds and lands. It is weaker on meandering conversation, where nothing in the words marks a beginning or an end, and it has no idea what is funny.
The hook-word and section lists are English. In other languages the energy, speech-rate, number and punctuation signals still work; the word-based ones don't fire, so clips lean more on delivery than on phrasing.
There is no face detection and no reading of the picture at all. The choice is made from the audio and the transcript.
Why a scoring system and not an AI model?
Three reasons, in order. It runs in seconds on any Apple Silicon Mac, so picking moments never waits on anything. It needs nothing from the internet — the transcript stays on your disk, which is the whole point of tondi. And it can be written down on one page, so when it picks a clip you disagree with, you can see why.
The cost is that it can't be surprised by a great line that doesn't look like one. If you want a model's opinion of what will perform, the cloud clippers and some local ones offer a virality score; tondi doesn't.
Common questions
Can I change how tondi picks clips?
You can set how many clips you want, and the shortest and longest length, which changes which stretches are candidates. The weights themselves aren't adjustable in the app.
Why did it pick a boring moment?
Usually because the delivery was loud or fast there, or the stretch happened to contain numbers and a question. The scoring reads signals, not meaning. Lowering the clip count keeps only the strongest scores.
Does tondi show me the scores?
No. Clips are shown in the order they occur in the video, not by score, so you judge them on what they are.
Does it get better over time?
No — nothing learns from your videos, because nothing about your videos is collected. The rules change only when the app is updated, and the changelog says when.