Repurpose by retention spike, not by chapter

Chapter markers are a map of the outline you wrote. Rewatch spikes are a map of the outline the audience actually used. Only one of those has been validated, and it is usually not the chapter titled "Tip 3."

Most repurposing pipelines still cut at the headings: pull the section, vertical-crop it, caption it, ship it. That is how you end up with a neatly labelled 40-second clip of a section people skipped, while a throwaway demonstration at 7:14 (the local bump above the retention line) never leaves the long file. The spike is a pre-tested hook. The chapter is a hope.

Chapters are your outline; spikes are the audience

A retention curve plots how many people remain at each timestamp, sometimes against videos of similar length. Practitioners already have a name for the local bump above the surrounding line: a rewatch spike. People dragged the playhead back, or replayed a loop, because something on screen was worth seeing twice. A plateau means they stayed. A spike means they returned.

Chapter markers cluster where you changed topics. Spikes cluster where they found a replayable unit: a demo that resolves in a few seconds, a number that does not make sense on first hearing, a visual that has to be paused to be read. Those often sit in the cracks between chapters, in asides, in the bit you almost cut.

The rest of a decent retention analysis is about naming drop shapes so you can repair the long file. This page is about mining the bumps so you can make short ones.

A transcript-scored clipper will not find this on its own. Scoring paragraphs for "hookability" is a reasonable first pass on a podcast with no video graph, and the long-form to Shorts playbook is built that way. When you have a retention graph, the graph outranks the transcript. The audience has already voted.

How to read a rewatch spike

Open the graph on the original upload, not on a previous clip. You want the timestamps the long-form audience actually used.

A spike is local. It rises above the line on both sides, then returns. A higher level after a drop is a recovery, not a spike. Recoveries help you edit the long file. They are weaker clip sources, because they often need the preceding minute to mean anything.

Ignore the first 15–30 seconds for mining. Opens are a different regime: misclicks, wrong-audience impressions, feed churn. A bump at 0:08 is usually people scrubbing to see if they should stay.

Mark the peak, then find the self-contained window around it. The peak is often the payoff. Walk backward until the unit has a beginning, and forward until it has an end. If that window will not fit inside 30–55 seconds, the spike is not a clip. Practitioner write-ups put the sweet spot for repurposed shorts in that band; treat it as a working band, not a platform law.

Prefer spikes that are visual. A replayed chart, a pair of hands, a then-and-now hold, a UI walkthrough. A replayed sentence without a picture is a quote card. Ship it as a quote card.

Log every spike before you cut anything:

Spike	Peak timestamp	Window (in–out)	Why it is replayable	Stands alone?
1	7:14	6:58–7:41	Demo of the fold, paused twice	Yes, if you start on the hands
2	12:02	11:40–12:35	Number that contradicts the title	No: needs the setup from 11:10
3	18:22	18:04–18:50	Then/now on the same frame	Yes

Row 2 does not ship. Forcing it to ship is how chapter-thinking sneaks back in: you "add the setup" until you have Tip 3 again.

The spike-first clipping procedure

Work from the original source file. Recut vertical from that file so the short does not look like an import. A file that still carries another platform's watermark or UI chrome is the thing Instagram lists as ineligible for recommendations (visible watermarks), and the thing other surfaces also downrank. Platform-native versus repurposed is the adaptation layer. This procedure is the selection layer that adaptation should start from.

1. Export or screenshot the retention graph of the long file. Mark every local bump above the line after the 30-second mark.
2. Fill the spike table. Kill any window that does not stand alone. Kill any window that cannot land inside 30–55 seconds without the setup.
3. Open on the spike, not on the chapter title. The first frame of the short is the most replayed frame, or the frame two seconds before it if the peak needs a beat of orientation. Do not open on your face restating the section name. That is a chapter clip with a spike buried at 0:12, which is to say a clip that will lose its audience before the validated moment.
4. Cut the window on an EDL timeline. Reframe to 9:16 from the source, not from a previously exported 16:9 file if you can avoid it. Burn captions as captions, not as a restated outline. The 480p preview pass is free of a credit charge and carries a short per-user cooldown; use it to confirm the first frame is the spike frame and that the window still makes sense with the long-form context stripped. The final export is charged once.
5. Write packaging from the spike, not from the chapter. Title and cover should promise the replayable unit ("the fold at 2x speed," "the number that does not match the label"), not "Part 3 of the workshop."
6. Strip foreign chrome. No other app's watermark, no leftover 16:9 letterbox, no "watch the full video" lower-third that reads as an import.

Throughput, as a working figure rather than a guarantee: practitioner write-ups put one 30-minute long-form at 6–10 clips in about 90 minutes of this kind of work, and a steady week at 5–7 shorts across surfaces from 2–3 originals. You will get fewer if you are honest about "stands alone." That honesty is the whole method. Count spikes shipped, not sections covered.

The paired test that proves it

"Spikes are pre-validated" is a mechanism, not a result. Prove it on your files with a test you cannot peek into.

For each of three long-form videos, cut two shorts:

• Spike clip. Window from the table. First frame is the spike frame. 30–55 seconds.
• Chapter clip. The section your outline would have pulled from the same video. Same length band. Same caption style, hashtags and sound. The only variable is the in and out points, plus the first frame that those in-points force.

Then:

1. Pre-commit the metric. Average watch time, or Viewed vs. Swiped Away on Shorts, read at a 24-hour checkpoint. Not views.
2. Do not post them simultaneously. They will cannibalise. Same clock time, consecutive days. Flip which one goes first across the three pairs so posting order cannot explain the result.
3. Do not look until the checkpoint. Early data on short-form is mostly which file hit the feed first.
4. Call a winner per pair, then across pairs. One pair is the feed lottery. Three pairs in the same direction is the smallest sample worth acting on. If they split 2–1, run two more. The A/B protocol is the same one you should already be using for hooks: one variable, consecutive days, metric first, stop rule written down.

If spike clips win on the metric, stop cutting chapters except as a last resort when a video has no spike. If they do not win, look at the losing spike clips before you throw out the method. The usual miss is opening two seconds too late (payoff with no orientation) or two seconds too early (the chapter heading).

Do not average the two files into "repurposing works." The question is which selection rule works. Volume of assets is a separate problem. This is a source-window problem sitting underneath it.

FAQ

What if the long video has no rewatch spike?

Then it has no pre-validated short. Do not invent one at a chapter boundary to hit a quota. Repair the long file first (a slow bleed and a mid-video cliff are different edits) and wait for a later upload that actually produces a bump. A channel with no spikes is not generating replayable units. Clipping will not create them.

Can I use auto-chapters from the transcript as a starting map?

You can use them to navigate to a spike you already marked. You should not use them as the cut list. Auto-chapters follow sentences. Spikes follow playhead behaviour. When the two agree, fine. When they disagree, the graph wins. If you only have a transcript (audio-only, no retention graph), fall back to scored standalone beats and accept that you are guessing.

Does this apply to a video that is already short?

Barely. A 40-second original has no room for a chapter-versus-spike argument. Mine spikes on files long enough to have an outline: roughly eight minutes and up, where a viewer can choose a region to replay. Shorts-to-Shorts "repurposing" is recutting the open, not mining a graph.

Should the short send people to the long video?

Not in the first 20 seconds, and not instead of making the short a complete unit. A short that only works as a trailer fails the "stands alone" column. If you want a bridge, put it after the payoff. End-screen CTAs sit on the smallest remaining audience. The spike clip's job is to be the file people finish.