YouTube chapters from a Twitch VOD

You re-uploaded last night's four-hour stream to YouTube and it sits there as one unbroken block. The progress bar gives a viewer nowhere obvious to start, so the VOD mostly idles while the clips do the work. Chapters change the shape of that page, and YouTube is specific about how you write them.

What YouTube requires

Three requirements come straight from YouTube's help page on adding video chapters. Treat them as conditions the block has to satisfy before YouTube will build chapters from it.

• The first timestamp has to be 00:00. Not 00:04, not 00:11.
• You need at least three timestamps, in ascending order.
• Every chapter has to run at least 10 seconds.

Make sure that the first timestamp you list starts with 00:00. Your video should have at least three timestamps listed in ascending order. The minimum length for video chapters is 10 seconds.

YouTube Help, Add video chapters

The timestamps live in the video's description. YouTube can also generate chapters on its own, and the same help page is careful about that: "Not all videos are eligible for automatic chapters, and not all eligible videos will have automatic chapters." A manual list overrides whatever it produced — the page says the option "will override automatic video chapters." You can also switch automatic chapters off for one video on its details page, or channel-wide in YouTube Studio under Settings, Upload defaults, Advanced settings.

Why a re-uploaded stream needs them

A stream VOD is an awkward shape for YouTube. It runs long, it has no cold open, and the opening minutes are often you fixing OBS. Chapters make the progress bar readable: YouTube splits it into labelled segments and lists the titles under the player, so somebody who arrived from a clip can land on the game switch instead of the mic check. That is a navigation fix for a video that is mostly ambient, and it is the whole case for doing it.

Where ClipEZ's chapters come from

ClipEZ runs chapter detection as a stage in the analysis pass, after transcription. The call sees the transcript and the stream's length — segment start times and segment text, nothing more. No chat log, no waveform, and no video frames reach it. The prompt asks for a broad table of contents for one continuous stream: a new game, a new activity, a long break, a technical issue, the wrap-up. It also tells the model to merge neighbouring chapters that cover the same activity, and to skip a boundary when it is unsure.

Two implementation details are worth knowing. Every chapter start time is snapped to an exact transcript segment boundary, so a chapter begins where a sentence begins. And the target chapter count comes from stream length — roughly one per fifteen minutes, capped at twenty, handed to the model as a soft floor and ceiling rather than a quota. The prompt tells it to return fewer when the stream is homogeneous, so three hours of one game comes back with fewer chapters than a variety stream of the same length.

A long chapter can carry subchapters — a boss attempt, a technical issue inside a longer phase. Those are navigation beats inside ClipEZ. YouTube's chapter format is flat, so nothing nested survives a paste into a description.

Producing the description block

Here is the part that surprises people. The YouTube chapters export reads your cuts. Detected chapters are how you find the boundaries; the file is built from what you marked.

1. Mark a cut at each point you want to become a chapter. Jump to a detected chapter, confirm it is a real phase change, and mark there.
2. Flag those cuts for export so they appear in the export bay's clip list.
3. Tick YT chapters in the export bay and run the export.

ClipEZ writes one file, youtube-chapters.txt, into the output folder — by default a folder beside your source video and named after it. Cuts are sorted by start time before anything is written, so the order you marked them in does not matter. Each line is a timestamp and a title. The title is the clip title when you have generated clip metadata, otherwise the label you typed on the cut, otherwise a plain Chapter 3. Under an hour the timestamp renders as MM:SS; past an hour it picks up the hour field.

If the first cut does not start at the origin, ClipEZ prepends an 00:00 Intro line so the block satisfies YouTube's first requirement. A cut starting anywhere inside the first second already renders as 00:00, and in that case the Intro line is skipped — two 00:00 lines would break the set. A finished file looks like this:

00:00 Intro
40:37 His crit build suddenly deletes the entire room
58:03 Gets saved by a resurrection relic after barely surviving the puppet fight
1:07:11 Survives an impossible boss attack and kills him mid-monologue
1:09:20 Thinks he died, then discovers he actually won

Where this breaks

• Timestamps are positions in the source VOD. Trim the intro before uploading and every chapter shifts by whatever you cut, which puts the whole block out. Upload the VOD unchanged, or re-mark after trimming.
• YouTube derives a chapter's length from the gap to the next timestamp. The length of your cut has no bearing on it, so two cuts five seconds apart produce a five-second chapter and fall under YouTube's 10-second minimum.
• You need three timestamps. One marked cut plus the automatic Intro line only gets you two.
• Titles are pasted verbatim into a public description and surface on the progress-bar tooltip. Keep them short.
• Chapters are not available on every video. YouTube's help page states that if the channel has any active strikes, or if the content may be inappropriate to some viewers, the feature won't be available.

What ClipEZ does not do here

It does not upload, write your description, or touch your channel. The export is a text file on your disk. You open the video in YouTube Studio, paste the block into the description, and save. If automatic chapters were on for that video, your pasted list replaces them.

Chapter detection saves you the pass where you scrub a four-hour timeline trying to remember when you switched games. The export saves you the formatting. Which phases deserve a line is still your call, and that is the part you were always going to judge better than a model reading a transcript.