Skip to main content

​
AI Clip Finder

The Clip Finder node uses AI to analyze a long video and identify the most viral segments for extraction.

​
Inputs

​
Outputs

The output is an array of video clips. This can be connected to a Split node for per-clip processing, or directly to batch-aware nodes.

​
Available engines

​
How it works

  1. Receives a long video from the input handle
  2. Transcribes audio using the transcription engine (word-level, aggregated into ~5s segments)
  3. Sends the timestamped transcript to Gemini Flash for viral potential analysis
  4. Ranks clips by viralness — hook strength, shareability, controversy, emotional impact, trend alignment, quotable soundbites
  5. Snaps AI-returned timestamps to actual transcribed word boundaries for precision
  6. Renders static JPEG thumbnails (one frame per clip) in a 3×2 grid to the right of the main preview — six clips per page, with Prev / Next when there are more than six. Thumbnails load one-by-one as their source frames decode, so each card swaps from spinner to image independently.
  7. The main preview is empty until a clip is selected — click any thumbnail’s play control to load just that trim into the player (no full-video preload). Click the image or title row to toggle selection.
  8. User selects one or more clips (multi-select with Select All / Deselect All) to continue down the pipeline
  9. On Use these clips, each selected clip is extracted server-side into its own small faststart .mp4 (via the hyperclip-extract-audio Lambda’s clip mode). Downstream nodes (auto-captions, video-editor, final render) then play a tiny standalone file instead of seeking into the long source — which is what fixes “black player that never loads” on hour-long uploads where the source isn’t faststart-muxed. Extraction is best-effort; any clip that fails extraction falls back to a trim into the original URL so the flow always continues.

​
How clips are picked

The ranker is briefed as a short-form editor (TikTok / Reels / Shorts), not a summariser. It’s looking for moments that work posted alone with no other context — strong opening hook, self-contained narrative, clear payoff, quotable lines, audible energy. Filler / transition / mid-thought passages are explicitly down-ranked. Each clip gets a viralness_score (0–100):

​
Configuration options

For API runs, pass under model_overrides[step]: Clips target 40–60 seconds (50s is the sweet spot for short-form retention). Hard floor: 25 seconds — clips under that lack the setup → payoff arc that keeps viewers watching, so they’re dropped before extraction. Hard ceiling: 60 seconds — anything longer is trimmed to 60s. If transcription fails or the LLM returns nothing parseable, the executor falls back to a single 60 s clip from the start of the source video.

​
Credit cost

1 credit per analysis (Gemini Flash).

​
Tips

  • Feed this long-form content from Media Upload or TikTok Scraper
  • The AI looks for moments with high viral potential — hooks, hot takes, emotional peaks
  • Use the play button on a thumbnail to preview that clip in the main player; click the thumbnail or title to select/deselect it
  • When you have more than six clips, use Next / Prev under the grid to page through them
  • Use “Select All” to process every detected clip, or pick specific ones
  • Combine with Video Editor to fine-tune the extracted clips

​
Connecting downstream

The output is a Video[] (dashed handle). The standard shape is:
Each clip runs through Auto Captions and Video Editor individually, so you can shape per-clip captions and edits. Both nodes detect when they’re inside a Split loop and surface shortcut buttons:
  • Auto Captions → Apply to All Remaining Clips — finishes the current iteration’s captions and applies the same style to every later iteration without opening the design UI.
  • Video Editor → Skip Editing for All — finishes the current iteration with no edits and skips the editor for every later iteration.
So you can either go through every clip individually or set the style/skip once and let the rest auto-complete. See Connecting Nodes — Arrays vs. single items for the validator rules and other array shapes.

​
Example use cases

  • Extracting viral-worthy moments from podcast recordings
  • Finding the best hooks in a recorded presentation
  • Creating short-form clips from long-form content for TikTok/Reels/Shorts