Skip to main content

Tighten a talking-head video (derush edit)

Goal: take a raw recording — a talking-head video, interview, vlog or screen recording with narration — and get a tighter edit with long pauses, false starts and repeated takes removed. The result is a new MP4 (edited.mp4).

"Derushing" is the editor's word for going through raw footage ("rushes") and keeping only the good parts. In ReelBolt an AI editor chooses which parts to keep, and ReelBolt itself does the precise cutting.

Prerequisites for a derush edit​

  • The video uploaded to the project (Files tab). Up to 30 minutes and 2 GB per video.

  • A chat AI provider for the editor.

  • Strongly recommended: a transcription provider, so the editor knows what is being said and keeps whole sentences. An admin adds one under Admin, Inference Providers with capability Transcription. Options:

    • the local whisper service that ships with ReelBolt (free, runs on your server; kind OpenAICompatible, endpoint http://whisper:8000);
    • an Azure OpenAI or other OpenAI-compatible speech-to-text service.

    Anthropic (Claude), Gemini and DeepSeek providers cannot transcribe. Without any transcription provider the edit still works, but the editor only sees pauses and shot changes, not words.

The derush workflow​

Add the Highlight edit from footage template (video-derush-edit) from New workflow. Its steps:

#LabelStep typeAgent
1Analyze source videoVideoAnalyzeVideoTransform (not an AI)
2Decide which spans to keepAgentVideoStoryEditor
3Compile edited videoVideoCompileVideoTransform (not an AI)
4Review edit qualityReviewLoopVideoReviewAgent (back to step 2 when below 8, up to 3 rounds)

Step 1 — open it and select your video as the source. Its settings then look like:

{"version": 1, "source": {"kind": "ProjectFile", "projectFileId": "PUT-THE-FILE-ID-HERE"}}

Speech-to-text is used automatically when a provider exists ("transcription": "Optional"). Use "transcription": "Required" to make the run fail clearly instead of editing without words, and "language": "en" (or your language) if detection guesses wrong.

Step 3 — as shipped:

{"version": 1, "decision": {"from": "Previous"}, "analysisStepOrder": 1,
 "transitionPolicy": "Auto", "programFadeInMs": 500, "programFadeOutMs": 800,
 "programAudioFadeInMs": 300, "programAudioFadeOutMs": 900, "minSegmentMs": 800}

This makes soft automatic joins between cuts, fades the start and the end, and drops kept pieces shorter than 0.8 seconds.

How to run a derush edit​

  1. Optional but useful: when you click Run, write a brief in the confirmation box to tell the editor what you want, for example "Cut to about 90 seconds, keep the story about the launch, drop the off-topic tangent near the end".
  2. Run it. Step 1 takes roughly as long as the video plus speech-to-text time; step 2 is one AI decision; step 3 encodes the video.
  3. Watch the result in step 3's Output Video, or in the project's Renders tab. The edit is also added to the project's Files.

What to check after a derush edit​

  • Step 1 result: meta.transcription.applied should be true if you expected words. If it says degraded, speech-to-text failed or no provider was found. If it shows droppedNoSpeech, that many "words" were thrown away because there was no speech under them: speech-to-text tends to invent lines such as "Thank you for watching." on music-only or silent clips, and ReelBolt discards them so the editor never mistakes them for dialogue. Clips with no sound at all are not transcribed.
  • Step 2 result: the list of kept spans (which ids were kept) and the editor's explanation of why.
  • Step 3 result: how long the result is, and the Open in Editor timeline to see each cut.
  • Step 4 result: the review score and comments. A low score sends the run back to step 2 with the comments.

Variation: a panel of editors (edit room)​

The Highlight edit, decided by an editing team template (video-derush-edit-room) replaces step 2 with an EditRoom step: three AI editors (pacing, story, craft) discuss the footage and a director writes the final decision. It is slower and uses more AI calls, but tends to make more considered choices on long or story-driven footage. If the panel fails, a single editor decides instead ("fallbackToSoloEditor": true).

{"version": 1, "view": {"from": "Previous"}}

You can give the room your own seats, for example:

{"version": 1, "view": {"from": "Previous"},
 "seats": [
   {"name": "BrandEditor", "persona": "argues for keeping every product mention and the call to action"},
   {"name": "PacingEditor", "persona": "argues for rhythm and removing dead air"}
 ]}

The room's discussion is kept on the step's result for reading; only the director's final decision counts.

Other variations of the derush edit​

  • Several takes or angles: in step 1 use "sources": [ ... ] with several videos instead of source. All clips are analysed together and the editor can pick the best parts from any of them.
  • Open on the best shot, not the first one: by default the parts of one video play in the order they were filmed. Set Order of the clips to In the editor's order in step 3 ("keepOrder": "AsListed") and say what you want in the brief, for example "open on the moment the product launches, then tell how we got there". The editor may then move a later moment to the front. It never shows the same moment twice. This needs the normal (re-encoded) cut.
  • Use only part of a video: in step 1, fill in Use from and Use until (seconds) on a clip to analyse and edit only that stretch of it ("inSec": 12, "outSec": 48 on the source). Everything outside it is left out of the edit.
  • Keep part of a long shot: set Split shots longer than in step 1 (for example 4 seconds, "maxShotSec": 4). Long shots are then cut into shorter parts, at a calm moment when possible, and the editor can keep only the good part of a long shot.
  • Plain cuts: set "transitionPolicy": "Off" in step 3 (and the fades to 0).
  • Fast, lossless cut: "mode": "StreamCopy" with "allowKeyframeSnapping": true in step 3. Cuts land on the nearest keyframe (less precise) and no extras (graphics, music...) are possible.
  • Describe shots with AI: "vision": "Optional" in step 1 adds a short AI description of what is in each shot (needs a Vision provider; costs one AI call per described shot, at most maxCaptionedShots, default 50).
  • Add extras: graphics, music, sound effects, colour grade, b-roll and voiceover each have their own recipe in this section.

Pitfalls of a derush edit​

  • Run fails at step 1 with SOURCE_UNRESOLVED: no video selected in step 1.
  • EMPTY_KEEP or EXPECT_FAILED at step 3: the editor kept nothing, or less than 15% of the footage. Give clearer instructions, or lower expect.minRetainedRatio in step 3 if a very short cut is intended.
  • Sentences cut mid-word: usually no transcription. Add a transcription provider.
  • Re-runs give the same edit: the template sets the editor to never reuse, but if you built your own workflow, set the editor step's cacheMode to "Never".
  • If you add steps between the editor and the compile step, change step 3's decision from {"from": "Previous"} to {"from": "Step", "stepOrder": 2} — "Previous" would then point at the wrong step.