How to turn a Twitch VOD into clips without rewatching it

You streamed for four hours last night. You know there were three or four moments worth posting — the clutch, whatever chat lost it over, the thing you said about the patch notes somewhere in hour two. This morning there is a multi-gigabyte file on your drive and no index into it except your memory.

Here is what the manual route actually costs, and what changes when the stream itself becomes searchable — including the parts of ClipEZ that are rough.

What the manual route costs

Route one: Twitch's clip editor

This is the path of least setup. Open the VOD on twitch.tv, drag the seek bar to roughly where you think the moment was, hit the clip button, trim it in the clip editor, publish.

Two things make it slow. The first is the seek bar. Four hours is 14,400 seconds compressed into one strip, so a single pixel of travel covers a long span of stream. You cannot land the mouse on a moment — you land in a neighbourhood and then hunt inside it, and every wrong guess costs a buffer and a reseek.

The second is the length ceiling. Twitch's own clip API documentation puts the editable range at "from 5 seconds in length to 60 seconds in length." That is fine for a headshot. A two-minute bit — a story, a rant, a bad take with a good ending — does not fit, so you cut it in half or drop it.

And the clip lives on Twitch. Getting the file into DaVinci Resolve or CapCut is a separate step, once per clip.

Route two: your OBS recording

If you record locally, the workflow looks like this:

1. Find last night's recording in your OBS output folder — an .mkv or .mp4 running many gigabytes.
2. Drop it on a timeline and wait while your editor indexes or conforms it.
3. Scrub at speed, hunting for a reaction you already lived through once.
4. When memory fails, open the VOD in a second window and scroll chat replay looking for a wall of emotes. Write the timestamp down. Go back to the editor and type it into the timecode field.
5. Mark in, mark out, name the range. Repeat for every moment.
6. Render each one.

Steps three and four are the evening. Everything else is minutes.

The transcript-and-timeline route

ClipEZ is a desktop app that sits upstream of your editor. It pulls the VOD, transcribes it, and gives you a transcript wired to a timeline and a player. You search text, jump to it, mark a clip, drag the edges, export. Here is the flow, step by step.

1. Paste the VOD link

Import takes a Twitch VOD URL like twitch.tv/videos/12345678, a YouTube link (watch, live, shorts, or a youtu.be short link), or a local file. The open dialog accepts mp4, mkv, mov, webm, avi, flv, ts and m4v, so an OBS recording works as it came off disk.

You do not need to connect your Twitch account to paste a Twitch link — linking is only for browsing your own VOD list inside the app. YouTube is the exception: ClipEZ checks through YouTube's API that the link is a past livestream rather than a regular upload, and that check needs your YouTube account linked. Audio is fetched before the video, so transcription starts while the video is still downloading.

2. Let the analysis run

The status bar names each stage as it goes: Probing video, Downloading audio, Transcribing, Polishing transcript, Segmenting, Building waveform, Detecting audio spikes, Fetching chat, Detecting chat bursts, Finding highlights with AI, Building chapters. How long it takes depends mostly on how fast the VOD downloads, which nobody controls, so there are no time promises here.

Worth knowing what leaves your machine on the managed path. Audio is extracted to 16 kHz mono AAC and uploaded in five-minute chunks — tens of megabytes per hour of stream. The transcript text goes to the model. The video file itself is never uploaded during analysis.

3. Search the transcript

This is the part that replaces the scrubbing. Cmd+F (Ctrl+F on Windows) opens a search bar at the top of the transcript panel. Three characters minimum. Enter walks forward through matches, Shift+Enter walks back, the up and down arrows do the same, and a counter shows which match you are on out of how many. Esc closes it. Click any transcript line and the player jumps there and starts playing.

So "where did I talk about the patch notes" becomes typing patch and pressing Enter twice. When a moment has no words attached — a clutch play with nothing said out loud — the chat-burst markers on the timeline are where to look instead.

4. Mark the clip

Three ways in:

• Drag down a run of transcript lines. The selection takes the first line's start and the last line's end and lands on the timeline as a range.
• Drag on the timeline to draw a selection, or set its edges from the playhead with I and O. Enter turns the selection into a clip.
• Right-click a block on the timeline and choose Add clip, or click the scissors on a row in the Highlights panel.

Marked clips get a CLIP badge on the timeline and appear in the Reel list in the export panel. Esc clears a selection you decided against.

5. Drag the boundaries

Every clip has drag handles on its in and out points, so tightening a cut is a mouse gesture. S splits the clip under the playhead. Delete removes the selected highlight, or the clip under the playhead. Cmd+Z and Cmd+Shift+Z step back and forward through the edit history.

Playback is what you would expect from an NLE. Space or K plays and pauses. J and L jump back and forward five seconds, as do the arrow keys; hold Shift and the arrows move ten. Shift+. steps the playback rate up through 2×, 3× and 5×, and Shift+, steps it back down. Scroll zooms the timeline, Shift+scroll pans, 0 fits the full duration, and the square brackets jump between highlights.

6. What the AI suggestions actually are

The Highlights panel lists suggestions with a score. Turn one into a clip with the scissors, dismiss it, or ignore the panel — the transcript and timeline work the same with suggestions switched off.

What sits behind them: the model receives the stream duration and one line per transcript segment — start time, duration, text. That is the whole input. No chat messages, no waveform, no audio levels. Suggested windows are then snapped to transcript segment boundaries, clamped between 10 seconds and 3 minutes, and de-duplicated where they overlap.

Audio spikes and chat bursts are computed locally on your machine and drawn as their own timeline markers. They never reach the model. They are there for you to read, and they are the second opinion when the transcript is quiet.

7. Export

Pick an output folder, tick the formats you want, pick a video quality, hit Export. Everything lands flat in that one folder — there is no per-export subfolder, so make a folder per stream unless you enjoy sorting afterwards.

Finishing in your editor

One thing to understand before you judge the cuts. The default export quality is Original, which stream-copies the video with no re-encode — fast, and no generation loss. The trade is that a stream copy can only start on a keyframe, so a clip may begin up to one keyframe interval early; on Twitch sources that is usually a couple of seconds. If you need the exact frame, pick a quality that re-encodes. Re-encodes use VideoToolbox on macOS and NVENC on NVIDIA Windows machines, with a software encoder as the fallback.

DaVinci Resolve and Final Cut Pro

Import timeline.fcpxml and every clip lands on one timeline in order. It is FCPXML version 1.10, and the media is referenced by absolute file path, so leave the source VOD where it is — move it and you will be relinking. The sequence is written at 1920×1080 and 30 fps whatever you streamed at, so check your project's frame rate after importing.

CapCut

The CapCut export writes a capcut-draft.capcut folder, and it is genuinely experimental. CapCut's draft format is not publicly specified, so the fields follow community-documented shapes. Installing it is manual: close CapCut, copy the whole folder into CapCut's drafts directory, relaunch. The README inside the folder has the exact path for macOS and Windows. If CapCut refuses to open the draft, re-saving it once from any CapCut version usually normalises it. The source video is referenced by absolute path here too, so re-link inside CapCut if you move it.

Straight to TikTok or Shorts

For clips going out as they are, MP4 plus the captions .txt covers it: one video file, one text file holding a title, a caption and hashtags. That text comes from the managed metadata step — generate it before exporting, or the file will say no caption was generated. For vertical, open a highlight in the short editor; it stacks the camera crop above the gameplay crop in a 9:16 frame and can burn the captions into the picture.

Before you try it

Constraints worth knowing up front:

• Videos have to be at least five minutes long. A 90-second recording gets rejected.
• macOS builds are Apple Silicon only. Windows builds are x64.
• ClipEZ is a paid desktop app — analysis and export both need an active subscription, and there is no free tier.
• Kick support is planned. Twitch, YouTube and local files are what work today.
• Export writes everything flat into the folder you choose.

None of this makes the clips for you. It removes the part where you scrub a four-hour file looking for something you already remember saying, which is the part that was eating the evening.