A transparent path from source video to publish-ready clips.

A new job begins with either one public direct MP4 URL or a file uploaded from the creator’s device. The inputs are mutually exclusive, and the video is the only required field. Clip duration, aspect ratio, resolution, caption style, language, and platform targets can use sensible defaults instead of turning the first screen into a professional editor setup form.

Device files move from the browser directly into private object storage through scoped multipart upload URLs. After the source is registered, the browser opens a route such as `/jobs/{id}`. That route owns progress, errors, selected moments, finished clips, and later downloads, so a refresh or return visit does not restart the work.

• Direct MP4 URL or device upload
• One required input
• Refresh-safe job route

FFmpeg extracts compact analysis-audio chunks so transcription can focus on spoken content without sending the complete video through every understanding step. Cloudflare Workers AI converts those chunks into timestamped English text, preserving the order and approximate boundaries of each phrase.

The transcript reveals where a premise begins, how an example develops, and when the central point resolves. It is a decision aid rather than a replacement for the source video. The current workflow does not add speaker diarization or pretend that transcript analysis can understand every visual cut, graphic, or demonstration.

• Compressed analysis audio
• Timestamped phrases
• English-first transcription

Candidate moments are evaluated for an understandable opening, enough context, specificity, standalone usefulness, and a natural ending. A sharp sentence that depends on missing setup should rank below a slightly longer passage that a new viewer can actually follow. Candidates are also compared as a set so several clips do not repeat the same lesson.

Source length sets a ceiling, not a quota: up to three clips for sources of three minutes or less, five up to ten minutes, eight up to thirty minutes, and twelve for longer sources. Confidence still decides the final count. Returning two strong clips is better than padding a job with weak variations.

• Complete setup and payoff
• Distinct themes
• Confidence before quantity

For every selected range, FFmpeg cuts the original source rather than reconstructing media from the transcript. The image is composed on a vertical canvas and the corresponding source audio is trimmed at the same boundaries, normalized, and kept synchronized with the picture.

The workflow deliberately excludes generated voice-over, background music, translation, and synthetic replacement audio. Because multimodal frame analysis is outside the current release, creators should inspect wide interviews, screen recordings, camera changes, and moving subjects in the rendered preview instead of assuming every automatic crop is correct.

• Original-source cut
• Vertical composition
• Synchronized retained audio

Clean-verbatim editing removes distracting filler, false starts, and obvious transcript artifacts while preserving the speaker’s claim, tone, order, and meaningful uncertainty. The text is grouped into stable phrases that can be read on a phone and placed where common feed controls are less likely to cover it.

Captions are burned into the downloadable MP4 and remain available as text in the job workspace. Review should cover names, numbers, timing, contrast, line breaks, safe-area placement, and agreement with the audible sentence. Caption polish must never turn a tentative statement into a stronger claim.

• Meaning-preserving wording
• Mobile-safe phrase groups
• Burned-in and copyable text

When rendering finishes, the same job route becomes a results workspace with inline 9:16 video players. Each result includes a suggested title, duration, highlight score, transcript or caption text, selection context, and a download action. The creator evaluates the complete edit instead of trusting a model score in isolation.

A final review starts from the first frame with no memory of the full source. The clip should orient the viewer quickly, preserve the important visual, keep captions and source audio synchronized, and end after the point lands. Any clip that fails that test should be revised or discarded before publication.

• Inline finished-video review
• Transcript and audio comparison
• Individual MP4 downloads