← Back to pipelines
Highlights
Read a transcript, return N editorial picks — the most quotable moments for clipping, chapter generation, or social posts.
ModelClaude Sonnet 4.6
10
1 – 20
Recipes using this pipeline
Highlights — Editorial Moment Picker
Drop in a transcript (SRT or plain text) and get back a ranked JSON list of the most quotable, hook-worthy moments. Claude reads the full text, applies your editorial directive, and returns each pick with a one-sentence context note and a target clip duration.
What you can do with it
- Feed the clip-factory recipe — the auto path in clip-factory runs highlights automatically between transcription and the per-clip trim loop, so you skip the manual clips.json step entirely
- Build chapter markers — use the picks to add YouTube chapter timestamps or podcast show-notes sections
- Generate social posts — each pick's context note is a ready-to-edit tweet or caption draft
- Steer the picker — pass a style hint like "the funniest moments" or "the strongest technical arguments" to get topic-specific cuts
How it works
- Transcribe your video or audio — run the Transcription pipeline to get an SRT asset in your library
- Attach the transcript and set the number of moments you want (default 5, max 20)
- Optionally add a style steer — plain English describing the kind of moments you want
- Download highlights.json — a JSON array of picks, each with a context note and desired_seconds
Frequently Asked Questions
What transcript formats are supported?
SRT (the standard subtitle format with timestamps) or plain text. SRT is preferred because the timestamps let the picker anchor each moment precisely. Plain text works too — the picker returns moments by content rather than timestamp position.
How does the style steer work?
Write a plain-English directive — 'the funniest moments', 'the strongest arguments', 'the most controversial takes', 'the emotional peaks'. Empty uses the built-in default: the most quotable, hook-worthy moments with strong opinions, surprising facts, vivid stories, and clear arguments.
How many moments should I ask for?
5 is a good default for a 30–60 minute recording. For a 10-minute video, 2–3 is usually right. For a multi-hour conference talk, 8–10. The picker clamps to however many distinct moments exist — a 5-minute clip with one main point will return 1 pick even if you ask for 5.
How does this work with the clip-factory recipe?
The clip-factory recipe runs highlights automatically — you don't need to call this pipeline separately unless you want to inspect or edit the picks before trimming. Pass --highlights-count and --highlights-style to the recipe to tune it.
How much does it cost?
1 credit flat, regardless of transcript length or number of picks. The cost is dominated by the LLM call, not input size.
