You finished a six hour stream. Somewhere in it are three or four moments worth cutting into clips, and you have a rough memory of two of them.
So you open the VOD and start dragging the scrubber. Six hours at 2x is three hours of watching yourself, and the moment you remember always turns out to be forty minutes from where you thought.
Almost nobody does this consistently. That is why most channels clip roughly nothing, and the clips that do exist were made by viewers who happened to be present.
The signal is already in the VOD
Chat reacted at the time. Every good moment in that six hours has a chat spike sitting on top of it, timestamped to the minute, recorded while it happened.
That is the whole idea. You do not need to remember where the moments were or watch the VOD again. You need the list of minutes where chat did something unusual, and then you need to look only at those.
For a typical stream that turns three hours of scrubbing into checking eight or nine timestamps.
Why the obvious alternatives do not work
Memory is worse than people think. You remember the moments you were proud of. Chat responds to different things, frequently to something going wrong.
The viewer graph is too slow. People leave gradually, so a dip points at something that happened several minutes earlier, and a genuinely great ninety seconds does not move it at all. Chat reacts inside the same minute. I went into why in more detail in what messages per minute measures.
Clipping live means clipping while playing. Most people are busy doing the thing worth clipping.
What counts as a spike
The rule needs to hold for a channel doing 300 messages a minute and one doing 4. A fixed threshold cannot do that, so the comparison has to be against the stream's own baseline.
We flag a minute when it beats all three of these at once:
- the stream's own average plus two standard deviations
- twice the stream's average
- a floor of 10 messages in the minute
The first two catch the shape of a spike. The floor stops a quiet channel going from 1 message to 4 from registering as an event, which is technically a 4x spike and obviously nothing.
Across the archive that has flagged 737 moments so far. The median runs 3 times the stream's normal pace and the largest hit 30 times.
Then each moment gets sorted by what chat was doing in it:
| Type | What triggered it |
|---|---|
| raid | raid keywords in the minute |
| hype | hype keywords in the minute |
| emote spam | laugh and pog emotes above 30 percent of the minute's messages |
| activity spike | everything else, chat simply went fast |
Most land in that last bucket, which is fine. The label is a convenience. The ranking is the part that does the work.
The detail that will save you a re-edit
Chat lags the moment. Someone has to see the thing, react, type it and hit enter, so the spike sits a few seconds behind the event that caused it, and further behind on a fast chat where the message has to queue.
So start the clip earlier than the flagged timestamp. Twenty to thirty seconds before is usually right. If you cut exactly at the spike you will catch the reaction and miss what people were reacting to, which is the single most common way this workflow goes wrong.
Where it does not help
Being straight about the limits, because they are real.
Quiet games produce quiet chat. A tense single-player run can have the best moment of the week land in near silence while everyone watches. Chat reaction finds what chat reacted to, and that is not the same as what was good.
Very small chats are noisy in the statistical sense. With 4 messages a minute as your baseline, one friend arriving looks like an event. The floor of 10 helps and does not fully solve it.
Slow mode flattens peaks by design. If you run slow mode during the parts of stream most likely to pop off, the tool is measuring the cap rather than the crowd.
And it only sees streams that were being recorded at the time. Chat is not recoverable after the fact, from us or anyone. It has to be captured live.
Doing it
Clip radar lists the flagged moments for a channel with timestamps, and the timestamps seek the VOD directly, so you can check a moment in a couple of seconds and move on.
If your channel is not recorded yet, add it. Recording starts within a few minutes and the moments accumulate from your next stream onward. Nothing before that point exists to analyse, which is the honest and slightly frustrating shape of this whole problem.