Highlight And Teaser Moments From Active Sequence
This workflow analyzes the currently active Premiere Pro sequence and marks short, self-contained moments that could work as teasers, highlights, podcast cold-open lines, social cutdowns, or review points.
Instead of searching the raw recordings manually, you ask the agent to understand the content actually used in the edit, find moments that can stand on their own, and add duration markers over the exact timeline ranges that make those moments useful.
New to Automation Agent?
This recipe uses Automation Agent for Adobe Premiere through the agent workflow via MCP. If you are here for the workflow solution first, that is fine. Use those pages for the product overview and setup path.
Video Walkthrough
When This Is Useful
Use this workflow when:
- you want to find strong teaser or highlight moments in an edited sequence
- you need podcast cold-open candidates made from sentences spoken later in the episode
- you want a review layer of candidate moments directly on the timeline
- you want each candidate to have a score and explanation, not just a timestamp
- you want the agent to inspect only the sequence content that viewers actually hear and see
Typical examples:
- podcast interviews where the opener should include two or three strong lines from later in the episode
- long interviews where the best standalone quotes should be surfaced for social clips
- tutorials or explainers where the most useful payoff moments should be marked
- presentations, demos, or product videos where a few moments should become promotional cutdowns
- documentary or event edits where a reviewer needs a first pass of emotionally or informationally strong moments
What You Need First
Before you run the prompt:
- Open the target sequence and make sure it is the active sequence in Premiere.
- Make sure Automation Agent is connected through your MCP workflow.
- Make sure the source clips used in the sequence have transcripts where possible.
- Decide whether you want the default number of candidates or a specific target count.
- If the sequence is visually driven, expect the agent to inspect a small number of targeted still frames where transcript context is not enough.
Recommended Execution Permissions
This workflow only needs to write sequence markers. To protect the rest of your edit, open the Execution Permissions tab before running it, enable Restrict write access in the Project section, then use + Add and choose Active Sequence Markers.
This adds write access for ::SEQ\ACTIVE::MARKER. The agent can create highlight and teaser markers on the sequence that is active for this run, but it cannot move clips, change track items, rename project items, or otherwise modify the project unless you add more write permissions.
Copy And Paste Prompt
Paste this prompt into your MCP client exactly as-is:
Using Automation Agent in Premiere Pro, analyze the currently active Premiere Pro sequence and create duration markers for strong highlight, teaser, and standalone pull-quote moments.
The goal is not to create chapters. The goal is to find short timeline ranges that could work as:
- teaser moments,
- highlight moments,
- podcast cold-open lines,
- social cutdown candidates,
- or concise overview moments that make a viewer curious.
Follow these steps carefully:
1. Analyze the active sequence timeline.
* Inspect all clips in the currently active sequence.
* For each clip, obtain the available transcript of its source media.
* Determine which part of the source transcript is actually audible in the sequence by comparing transcript timing with the clip's in/out range as used in the timeline.
* Ignore transcript sections that belong to unused parts of the source clip.
* Do not analyze the full source clip transcript blindly. Only use transcript portions that are actually present in the edited sequence.
2. Build a timeline-aware transcript.
* Combine the audible transcript parts from all clips in the order they appear in the sequence.
* Preserve the temporal context: keep track of when each spoken sentence or short passage happens in the timeline.
* If multiple clips overlap, focus on the speech that is actually relevant and audible in the final sequence.
* Keep the relationship between transcript text and sequence timecode clear, so duration markers can later cover the exact spoken or visual moment.
3. Understand the content and editorial purpose.
* Read the timeline-aware transcript as the actual spoken content of the edited video.
* Look for sentences, exchanges, explanations, reveals, claims, emotional beats, jokes, conflicts, insights, or summaries that can stand alone without too much missing context.
* Prefer moments that create curiosity, communicate a strong idea, or make the viewer want to keep watching.
* For podcast-style material, especially look for one- to three-sentence moments that could be placed at the beginning as a cold open.
* Avoid selecting moments that only make sense after a long setup unless the selected range includes the needed setup.
4. Use visual analysis only when helpful.
* Start with transcript-based analysis.
* If the transcript is sufficient, do not export unnecessary still images.
* If a candidate moment depends on visuals, or the transcript is incomplete or ambiguous, export selected still images from the sequence for visual inspection.
* Visual analysis is especially useful for:
* screen recordings,
* product demos,
* visual reveals,
* before/after examples,
* slides or presentations,
* montage sections,
* long transcript gaps,
* or visually important reactions.
* Do not export frames continuously or at fixed short intervals by default.
* Export still images only at strategically useful positions, such as inside a likely candidate range or around an unclear visual reveal.
* Use the visual information to refine the candidate range and explanation, not to create noise.
5. Select candidate moments.
* Choose only moments that are genuinely useful as highlights or teasers.
* A candidate may be a single sentence, a short exchange, a visual reveal, or a compact passage.
* Each candidate must have a clear start and end time in the active sequence.
* The marker range must cover the full usable moment, not just its start.
* Prefer tight ranges, but include enough context for the moment to make sense.
* Avoid overlapping candidates unless both versions are meaningfully different.
* Do not create candidates merely because there is a cut, pause, or topic change.
* If no target count was given by the user, choose a sensible number based on sequence length and quality:
* under 15 minutes: usually 3 to 6 candidates,
* 15 to 45 minutes: usually 5 to 10 candidates,
* 45 to 120 minutes: usually 8 to 16 candidates,
* over 120 minutes: usually 12 to 20 candidates.
* Quality beats quantity. If only a few moments are strong, create only those.
* If there are many strong moments, prefer the strongest candidates and mention that more could be extracted in a second pass.
6. Score and label each candidate.
* Give every final candidate a score from 1 to 100.
* Use the score to represent how strong the moment is as a teaser or highlight, not how important it is to the full video.
* Score higher when the moment:
* works without much setup,
* creates curiosity,
* contains a strong quote or reveal,
* has emotional or informational punch,
* is easy to reuse in an opener or cutdown,
* and has a clean start and end.
* Assign one primary type: Teaser, Highlight, Pull Quote, Overview, Reveal, Reaction, or Social Cutdown.
* Write a short title that describes the moment in plain language.
7. Add duration markers to the Premiere Pro sequence.
* For every final candidate, create a sequence marker on the active sequence.
* The marker must be a sequence marker, not a clip marker.
* The marker must have both a start time and a duration that covers the complete candidate range.
* Use marker type COMMENT unless there is a specific reason to use another marker type.
* Use this marker name format:
`92 Teaser - Short moment title`
* Replace `92` with the actual score, replace `Teaser` with the primary type, and keep the title concise.
* Do not include timestamps in the marker name.
* Put the detailed reasoning in the marker comment using this format:
Score: 92/100
Type: Teaser
Use: Podcast cold open, social teaser
Quote or beat: "Short representative quote or visual beat"
Why: One or two sentences explaining why this moment was selected.
Notes: Any uncertainty, transcript limitation, visual dependency, or suggested handle adjustment.
* If the exact frame is uncertain, choose the tightest reliable range and mention the uncertainty in the comment.
8. Output the result in the chat.
* After creating the markers, provide a ranked list of the selected moments.
* For each moment, include start time, end time, score, type, marker name, and a short reason.
* Also briefly mention that the same moments were added as duration markers in Premiere Pro.
* If visual still images were used, briefly mention that selected frames were inspected.
* If some clips had missing or incomplete transcripts, mention this limitation briefly.
Important quality guidelines:
* Base the analysis primarily on transcript sections that are actually audible in the edited sequence.
* Do not use unused parts of source clip transcripts.
* Create duration markers, not instant-only markers.
* Do not mark every interesting sentence. Mark the moments that would actually help an editor choose teasers or highlights.
* Keep marker ranges tight but usable.
* Use score and comments to make the review pass easy to scan.
* If the sequence contains very little spoken or visually interpretable content, create only the candidates that are justified by the available information.
* Before creating the final markers, reason through the candidate list carefully and remove weak or redundant moments.
Implementation Prompt For Repeatable Or Local Runs
If you want to run this workflow repeatedly, or with a local / limited-context MCP client such as LM Studio, use an implementation prompt instead of the high-level prompt above.
This version gives the agent concrete Automation Agent MCP and Runtime DSL patterns up front. That reduces the amount of discovery the agent has to redo on every run.
Copy implementation prompt
You are an MCP agent controlling Adobe Premiere Pro through Automation Agent Runtime DSL.
Analyze the currently active Premiere Pro sequence, identify strong teaser, highlight, pull-quote, overview, reveal, reaction, or social-cutdown moments from the edited sequence content, and add COMMENT duration markers to the active sequence.
This is not a chaptering workflow. Do not mark every topic change. Mark short timeline ranges that help an editor find reusable moments for teasers, cold opens, social cutdowns, or concise overview clips.
Workflow order:
1. Call `validateDsl` for the read-only active-sequence audio clip discovery probe.
2. If validation succeeds, call `runDslProgram` for that discovery probe.
3. Parse `CLIP|...` records from the result logs.
4. Determine distinct source `PROJECT_ITEM_REF` values used by audible timeline audio clips.
5. For each relevant source item:
- replace placeholders in the transcript export skeleton
- call `validateDsl`
- only if validation succeeds, call `runDslProgram`
- read and parse the exported transcript JSON file
6. Build a timeline-aware transcript from only audible edited sequence ranges.
7. Select and score final candidate moments.
8. Before marker writing, summarize the final candidate list in chat or internal working notes with start, end, score, type, marker name, and short reason. This is a sanity step, not an approval gate unless the user explicitly asks for approval.
9. Generate marker creation DSL from the marker skeleton.
10. Call `validateDsl` for the concrete marker creation DSL.
11. Only if validation succeeds, call `runDslProgram` for marker creation.
12. Call `validateDsl` for the marker read-back probe.
13. If validation succeeds, call `runDslProgram` for marker read-back.
14. Confirm created marker names, starts, and durations from `MARKER|...` records.
15. Output a ranked list of selected moments in chat.
Validation rule:
- Always call `validateDsl` before `runDslProgram`.
- Only call `runDslProgram` after validation succeeds.
- If validation fails, repair only the specific reported issue, then call `validateDsl` again.
- Do not run failed DSL.
Required read-only discovery:
Run the active-sequence audio clip discovery probe before any `prMarkerCreate`.
Do not guess project item refs, clip ranges, source in/out points, or whether transcript times map directly to sequence times.
Expected discovery output:
`CLIP|t=<audioTrackIndex>|i=<clipIndex>|name=<clipName>|start=<sequenceStartSeconds>|end=<sequenceEndSeconds>|in=<sourceInSeconds>|out=<sourceOutSeconds>|piRef=<projectItemRef>|piName=<projectItemName>`
Ready-to-run DSL: Active sequence audio clip discovery
```json
{
"version": 0,
"entry": "b1",
"blocks": {
"b1": {
"type": "prForEachTrack",
"inputs": {
"trackVar": "trackRef",
"indexVar": "trackIdx",
"kind": "AUDIO"
},
"body": [
{
"type": "prForEachTrackItem",
"inputs": {
"itemVar": "clipRef",
"indexVar": "clipIdx",
"track": { "var": "trackRef" },
"kind": "CLIP"
},
"body": [
{
"type": "esConsoleLog",
"inputs": {
"message": [
"CLIP|t=", { "var": "trackIdx" },
"|i=", { "var": "clipIdx" },
"|name=", {
"type": "prGetTrackItemProperty",
"inputs": {
"item": { "var": "clipRef" },
"property": "NAME"
}
},
"|start=", {
"type": "prGetTrackItemProperty",
"inputs": {
"item": { "var": "clipRef" },
"property": "START_SECONDS"
}
},
"|end=", {
"type": "prGetTrackItemProperty",
"inputs": {
"item": { "var": "clipRef" },
"property": "END_SECONDS"
}
},
"|in=", {
"type": "prGetTrackItemProperty",
"inputs": {
"item": { "var": "clipRef" },
"property": "IN_POINT_SECONDS"
}
},
"|out=", {
"type": "prGetTrackItemProperty",
"inputs": {
"item": { "var": "clipRef" },
"property": "OUT_POINT_SECONDS"
}
},
"|piRef=", {
"type": "prGetTrackItemProperty",
"inputs": {
"item": { "var": "clipRef" },
"property": "PROJECT_ITEM_REF"
}
},
"|piName=", {
"type": "prGetTrackItemProperty",
"inputs": {
"item": { "var": "clipRef" },
"property": "PROJECT_ITEM_NAME"
}
}
]
}
}
]
}
],
"next": null
}
}
}
```
Transcript export skeleton:
- This writes a local transcript JSON file.
- It does not modify the Premiere project.
- It is a skeleton, not ready-to-run DSL.
- Replace every placeholder, then call `validateDsl` before `runDslProgram`.
Target agent may adapt:
- `__PROJECT_ITEM_REF__`
- `__TRANSCRIPT_OUTPUT_PATH__`
Target agent must inspect instead of guessing:
- project item ref from `CLIP|...|piRef=...`
- a writable output path
- whether the clip using that source is actually audible in the sequence
Expected output:
`TRANSCRIPT_EXPORTED|path=<transcriptOutputPath>`
```json
{
"version": 0,
"entry": "b1",
"blocks": {
"b1": {
"type": "setVar",
"inputs": {
"name": "tx",
"value": {
"type": "prExportTranscriptJson",
"inputs": {
"item": "__PROJECT_ITEM_REF__"
}
}
},
"next": "b2"
},
"b2": {
"type": "writeTextFile",
"inputs": {
"path": "__TRANSCRIPT_OUTPUT_PATH__",
"text": { "var": "tx" },
"append": false,
"flushNow": true
},
"next": "b3"
},
"b3": {
"type": "esConsoleLog",
"inputs": {
"message": "TRANSCRIPT_EXPORTED|path=__TRANSCRIPT_OUTPUT_PATH__"
},
"next": null
}
}
}
```
Timeline-aware transcript rules:
- Use the transcript JSON exported from `prExportTranscriptJson`.
- Ignore words whose `text` is empty.
- Combine words into sentence-like passages until a word has `eos: true`.
- Use only source transcript words that intersect a timeline clip's source in/out range.
- Do not analyze unused source transcript sections.
Map source transcript time to sequence time:
```text
sequenceTime = clipStartSeconds + (sourceTranscriptTime - clipInPointSeconds)
```
Allowed assumption:
- If inspection shows one full-length audio clip with `start = 0`, `in = 0`, and `end = out`, source transcript times map directly to sequence times.
Must inspect:
- whether the full-length/direct mapping assumption is true
- whether there are multiple audio clips
- whether audio clips overlap
- whether a source transcript covers the clip's used source range
Candidate selection:
- Select only strong editorial moments.
- Quality beats quantity.
- If only a few moments are strong, create only those.
- Avoid selecting moments only because a cut, pause, or topic change occurred.
- Avoid duplicate or redundant candidates.
- Avoid unused source transcript content.
Default target count:
- under 15 minutes: usually 3 to 6 candidates
- 15 to 45 minutes: usually 5 to 10 candidates
- 45 to 120 minutes: usually 8 to 16 candidates
- over 120 minutes: usually 12 to 20 candidates
Candidate types:
- `Teaser`
- `Highlight`
- `Pull Quote`
- `Overview`
- `Reveal`
- `Reaction`
- `Social Cutdown`
Score each final candidate from 1 to 100.
Score higher when the moment:
- works without much setup
- creates curiosity
- contains a strong quote or reveal
- has emotional or informational punch
- is easy to reuse in an opener or cutdown
- has a clean start and end
Write behavior:
- `writeTextFile` creates or overwrites the local transcript JSON export path chosen by the target agent.
- `prMarkerCreate` adds COMMENT duration markers to the active sequence.
- The workflow adds sequence markers only.
- It does not create clip markers.
- It does not clear existing sequence markers.
- It does not delete, overwrite, rename, move, relink, or trim clips.
- It does not rename project items, bins, media, or sequences.
- Do not use marker deletion blocks in this workflow.
Marker creation skeleton:
- This is a skeleton, not ready-to-run DSL.
- Replace every placeholder and repeat the block pattern for every final candidate.
- Then call `validateDsl` before `runDslProgram`.
Marker owner:
`"::SEQ\\ACTIVE"`
Marker type:
`"COMMENT"`
Marker name format:
`<score> <Type> - <Short title>`
Marker comment format:
```text
Score: 92/100
Type: Teaser
Use: Podcast cold open, social teaser
Quote or beat: "Short representative quote or visual beat"
Why: One or two sentences explaining why this moment was selected.
Notes: Any uncertainty, transcript limitation, visual dependency, or suggested handle adjustment.
```
How to repeat the marker structure:
- Create one `prMarkerCreate` block per candidate.
- Use unique block ids such as `m1`, `m2`, `m3`.
- Set `entry` to the first marker block id.
- Set each block's `next` to the next marker block id.
- Set the final block's `next` to `null`.
- Do not include placeholder strings in the final DSL.
- `startSeconds` and `durationSeconds` must be numbers, not strings, in the final concrete DSL.
- Validate the final concrete DSL before running.
Skeleton for two markers:
```json
{
"version": 0,
"entry": "m1",
"blocks": {
"m1": {
"type": "prMarkerCreate",
"inputs": {
"owner": "::SEQ\\ACTIVE",
"name": "__SCORE_1__ __TYPE_1__ - __SHORT_TITLE_1__",
"type": "COMMENT",
"startSeconds": "__START_SECONDS_1__",
"durationSeconds": "__DURATION_SECONDS_1__",
"comments": "Score: __SCORE_1__/100\nType: __TYPE_1__\nUse: __USE_1__\nQuote or beat: \"__QUOTE_OR_BEAT_1__\"\nWhy: __WHY_1__\nNotes: __NOTES_1__"
},
"next": "m2"
},
"m2": {
"type": "prMarkerCreate",
"inputs": {
"owner": "::SEQ\\ACTIVE",
"name": "__SCORE_2__ __TYPE_2__ - __SHORT_TITLE_2__",
"type": "COMMENT",
"startSeconds": "__START_SECONDS_2__",
"durationSeconds": "__DURATION_SECONDS_2__",
"comments": "Score: __SCORE_2__/100\nType: __TYPE_2__\nUse: __USE_2__\nQuote or beat: \"__QUOTE_OR_BEAT_2__\"\nWhy: __WHY_2__\nNotes: __NOTES_2__"
},
"next": null
}
}
}
```
Representative validated concrete example with two markers:
```json
{
"version": 0,
"entry": "m1",
"blocks": {
"m1": {
"type": "prMarkerCreate",
"inputs": {
"owner": "::SEQ\\ACTIVE",
"name": "94 Reveal - The near-term AI danger",
"type": "COMMENT",
"startSeconds": 2942,
"durationSeconds": 83,
"comments": "Score: 94/100\nType: Reveal\nUse: Strong teaser, cold open, policy cutdown\nQuote or beat: \"autonomous weapons that could essentially take the decision to kill by themselves\"\nWhy: A concrete risk moment with strong policy stakes.\nNotes: Example only. Do not reuse these values unless this exact timeline and transcript mapping are confirmed."
},
"next": "m2"
},
"m2": {
"type": "prMarkerCreate",
"inputs": {
"owner": "::SEQ\\ACTIVE",
"name": "93 Teaser - AI expands what we are",
"type": "COMMENT",
"startSeconds": 3527,
"durationSeconds": 67,
"comments": "Score: 93/100\nType: Teaser\nUse: Podcast cold open, optimistic social teaser\nQuote or beat: \"Technology is expanding what we are.\"\nWhy: A concise thesis moment that works outside the full interview.\nNotes: Example only. Do not reuse these values unless this exact timeline and transcript mapping are confirmed."
},
"next": null
}
}
}
```
Ready-to-run DSL: marker read-back probe
Expected output:
`MARKER|i=<index>|name=<markerName>|start=<startSeconds>|duration=<durationSeconds>`
```json
{
"version": 0,
"entry": "b1",
"blocks": {
"b1": {
"type": "prForEachMarker",
"inputs": {
"owner": "::SEQ\\ACTIVE",
"markerVar": "markerRef",
"indexVar": "markerIndex"
},
"body": [
{
"type": "esConsoleLog",
"inputs": {
"message": [
"MARKER|i=", { "var": "markerIndex" },
"|name=", {
"type": "prGetMarkerProperty",
"inputs": {
"marker": { "var": "markerRef" },
"property": "NAME"
}
},
"|start=", {
"type": "prGetMarkerProperty",
"inputs": {
"marker": { "var": "markerRef" },
"property": "START_SECONDS"
}
},
"|duration=", {
"type": "prGetMarkerProperty",
"inputs": {
"marker": { "var": "markerRef" },
"property": "DURATION_SECONDS"
}
}
]
}
}
],
"next": null
}
}
}
```
Existing project state handling:
- Do not remove existing sequence markers.
- Do not overwrite existing sequence markers.
- New markers are appended.
- If existing markers have similar names, mention possible duplicates in the final summary.
- Do not edit, trim, move, rename, relink, or replace clips.
- Do not modify project items.
- Only use inspected `PROJECT_ITEM_REF` values for transcript export.
Missing transcript handling:
- Continue with clips that have usable transcripts.
- Mention missing transcript limitations.
- If no usable transcript remains, stop without writing markers.
- Do not infer content from unused source transcript or source filename.
Final chat output:
After marker creation and read-back, provide a ranked list.
For each moment include:
- start time
- end time
- score
- type
- marker name
- short reason
Also state:
- that the same moments were added as COMMENT duration markers in Premiere Pro
- whether transcript timing was direct or mapped through clip in/out
- whether any transcripts were missing or incomplete
- whether no visual stills were used
Final checklist:
- Active sequence inspected with read-only audio clip probe.
- All relevant audio clips have `start`, `end`, `in`, `out`, and `PROJECT_ITEM_REF`.
- Transcript export DSL validated for each source item.
- Transcript JSON exported and parsed.
- Timeline-aware transcript built from audible clip ranges only.
- Candidate list selected, scored, typed, and de-duplicated.
- Marker names follow `<score> <Type> - <Short title>`.
- Marker comments follow the required multiline format.
- Concrete marker creation DSL contains no placeholders.
- Concrete marker creation DSL validated successfully.
- Marker creation executed successfully.
- Marker read-back probe executed successfully.
- Final response mentions any transcript limitations.
- No clips, project items, sequences, bins, or existing markers were deleted, renamed, moved, or overwritten.
What The Agent Will Do
In practical terms, a good run of this workflow should:
- inspect the active sequence and identify the clips that contribute to the final edit
- gather source transcripts and map only the used portions into sequence time
- reconstruct a timeline-aware transcript for the edited video
- identify short moments that can work as teasers, highlights, or standalone pull quotes
- use targeted still-frame inspection only where transcript context is not enough
- choose a practical candidate count based on sequence length and moment quality
- create duration markers that cover the full candidate ranges
- label each marker with score, type, and a concise title
- put the detailed rationale and any uncertainty into marker comments
- return a ranked list of the marked moments in the chat
Expected Result In Premiere
The result should be:
- a set of duration markers on the active sequence
- one marker per selected teaser or highlight candidate
- marker names that are easy to scan in the Markers panel, such as
91 Teaser - The impossible tradeoff - marker comments that explain the score, use case, quote or beat, and reason for selection
- ranges that cover the complete usable moment rather than only a single timestamp
This workflow is meant to produce an editorial review layer. The agent should surface strong candidates, but the editor still decides which moments become the final opener, trailer, short, or cutdown.
Marker Naming And Comments
Use the marker name for fast scanning and the marker comment for editorial detail.
Recommended marker name:
92 Teaser - Short moment title
Recommended marker comment:
Score: 92/100
Type: Teaser
Use: Podcast cold open, social teaser
Quote or beat: "Short representative quote or visual beat"
Why: The moment works without much setup and creates a clear open question for the viewer.
Notes: Start could be extended by 0.5 seconds if the cut feels too abrupt.
This keeps the marker list readable while still preserving the reasoning behind each choice. Putting the score at the start of the name makes the strongest candidates visible immediately, while the comment remains the place for nuance.
Prompt Variants
You can adapt this workflow depending on the kind of highlights you want.
1. Podcast cold-open candidates only
If you are preparing a podcast intro, add this instruction:
Only select moments that could work as podcast cold-open lines before the episode starts. Prefer one- to three-sentence spoken moments that are understandable without setup, create curiosity, and do not spoil the entire discussion.
That is useful when:
- you want two or three strong lines for the beginning of an episode
- the final use is audio-first or dialogue-first
- you do not want broader visual highlights
2. Social cutdown candidates
If the goal is short-form clips, add this instruction:
Prefer moments that can become standalone 15- to 60-second social cutdowns. Include enough setup inside each marker range for the clip to work when separated from the full video.
That is useful when:
- the marked ranges may become reels, shorts, or trailers
- a single quote is too short without surrounding context
- you want fewer but more complete ranges
3. Tight pull quotes only
If you want the shortest reusable quotes, add this instruction:
Prefer tight pull-quote markers over longer highlight ranges. Each marker should usually cover one complete sentence or one compact exchange.
That is useful when:
- you are collecting text quotes, captions, or opener lines
- you want less context and faster review
- the sequence has many strong spoken moments
4. Give the agent a target count
If you already know how many candidates you want, add a target count:
Create exactly 8 final candidate markers unless there are not 8 moments that meet the quality bar. If fewer than 8 are justified, create fewer and explain why.
That is useful when:
- the sequence is long and you want a controlled first pass
- multiple editors will compare candidate lists
- you want a predictable amount of marker noise on the timeline
5. Rank for a specific audience or campaign
If the strongest moment depends on the intended viewer, give the agent that context:
Rank moments for an audience of first-time viewers who do not know the speaker. Prefer moments that communicate stakes, surprise, or a clear practical takeaway.
That is useful when:
- the best highlight depends on marketing strategy
- internal context would otherwise make weak moments look stronger than they are
- you want the score to reflect a particular use case
6. Require visual verification for visual reveals
If the sequence contains important visual moments, add this instruction:
For any candidate that depends on a visual reveal, product state, screen recording, slide, or reaction shot, inspect at least one still frame inside the proposed marker range before creating the marker.
That is useful when:
- the transcript alone cannot prove the moment works
- a visual punchline or reveal is more important than the spoken words
- marker comments should explain both the spoken and visual value
Related Workflows
This recipe is closely related to YouTube Chapter Markers From Active Sequence. Both workflows reconstruct the edited sequence from transcript portions and targeted visual inspection. The difference is the editorial goal:
- YouTube chapters mark section starts with instant markers.
- Highlight and teaser moments mark reusable ranges with duration markers, scores, and review comments.