The job splits into two halves that have almost nothing to do with each other. One is reformatting: taking a chosen moment and making it vertical, captioned, and the right length for each platform. That half is a solved editing task and dozens of tools do it. The other half is deciding which forty seconds out of a five hour stream deserve to exist as a short, and no format converter touches it. Everything below is about the second half, plus the one step that has to happen before either.
Get the file before it expires
Twitch does not keep VODs indefinitely, and how long yours stays up depends on your account type. Broadcast archiving also has to be switched on before the stream, not after. Anything you plan to clip later should be downloaded to your own storage as a file, because most of what you will use later, editors included, works on files rather than channel links.
Marking while you stream
The cheapest method costs nothing and happens live: hit the clip button, or trigger a marker, the moment something lands. The clip exists before the stream ends and the timestamp is exact.
Two limits come with it. You are judging mid-stream while doing something else, which means you catch the obvious moments and miss the quiet ones. And there is lag between the thing happening and you reacting, so markers tend to land late. Setting this up properly, including a hotkey you can hit without breaking flow, is covered in marking moments while you are still live.
Reading the signals the stream already produced
Chat is a rough highlight detector. A spike in message volume usually sits a few seconds after something worth watching, and the clips viewers made themselves are a shortlist somebody else built for free.
What this misses is everything the audience did not react to loudly: a good explanation, a clean play in a quiet moment, a line that reads better out of context than it did live. It also skews toward whatever chat finds funny, which is not always what performs on a platform where nobody knows you.
Automatic clipping tools
There is a whole category that takes the full VOD and returns finished vertical clips with captions. They score segments using some mix of transcript, audio energy, and pacing, then cut and reframe. For talking-heavy content, especially a podcast-shaped stream, the output is often usable with light trimming.
The tradeoff is that you get the tool's taste rather than yours, and its taste is tuned for speech. Gameplay moments where the meaning is on screen tend to be scored poorly, because the audio around a good play sounds like the audio around a bad one. Whether this category can do the whole job on its own is the starting question in the automated version of this workflow.
Searching the VOD by describing the moment
The other approach keeps the choosing with you and makes the finding fast. The VOD is uploaded to a system that indexes it once, and afterwards you search it by describing what you want in plain language.
In practice there are two different passes, and they answer different questions. A quick pass tells you which recordings contain something relevant at all, which is useful across a batch of files and nearly useless when you have one long VOD, since the answer is always yes. The narrower pass is the one that matters here: it returns short ranges with a sentence explaining why each matched, so you can scan the reasons before opening anything. It takes longer, the whole run has to finish before results appear, and it draws on more of your search allowance than the quick pass does. That is the tradeoff to expect. Vivu works this way on footage you have uploaded to a project: you describe the moment the way you would describe it to someone who watched the stream, open the ranges that come back, and export the original clip to take into your editor. It does not make the vertical cut, write captions, or pull a VOD from a link, so the download and the edit stay yours.
Searching by description is strongest exactly where the automatic tools are weakest, which is anything visual: a specific thing that appeared on screen, a reaction, a moment you remember by what it looked like rather than what was said.
When none of this is worth the effort
If you stream twice a month and already know the three moments worth clipping, open the VOD and cut them. Tooling pays off at the point where you cannot remember what happened, which arrives with volume rather than with ambition.
It is also worth checking whether shorts are the right output at all. Clips from a stream work when the moment stands on its own without the hour of context around it. A lot of good stream content does not, and cutting it shorter does not fix that, it just makes the lack of context more obvious.
The choice between the two real approaches comes down to who picks. If you want volume and are willing to accept someone else's judgment about what was good, the automatic tools are faster and always will be. If the moments you want are ones only you would recognize, you need to be able to ask for them specifically, and that means finding them yourself with something that understands the description. Most people who stream regularly end up wanting both, for different weeks. The single-destination version of this, when only YouTube matters, is laid out in the Shorts-only workflow.
FAQ
Can I paste a Twitch VOD link into an AI tool and get clips back?
Some tools accept a link and many do not. Anything that has to index or analyze the video usually needs the file itself, which means downloading the VOD first. Plan on having local copies of anything you intend to clip later, both because of that and because Twitch stops hosting VODs after a window that depends on your account type.
Should the same clip go to both TikTok and YouTube Shorts?
The cut can usually be the same, but the opening rarely should be. The first second does different work on each platform, and captions styled for one often look wrong on the other. Reuse the selected moment, re-do the hook and the text, and expect the same clip to perform differently in each place for reasons nobody outside the platform can explain.
Do I need a transcript to find clips in a stream?
Only for moments defined by what was said. A transcript is the fastest way to locate a specific line, a story you told, or an answer to a question from chat. It cannot find anything visual, which on a gameplay stream is most of what you want. If your streams are mainly talking, a transcript covers a lot of ground; if they are mainly playing, it will come up empty far more often than you expect.
How long should a clip from a stream be?
Long enough to stand alone without the stream around it, which is usually longer than the moment itself. Clips almost always need padding at the front for setup and at the end for the reaction, and a tightly-trimmed highlight tends to read as confusing rather than punchy. Cut the moment first, then add back whatever a stranger needs in order to follow it.