The Warlock Commandery Taroos the Warlock

Projects / Media / Video Work / Warlock Video Producer

Active build Updated 28 SEP 2026videoffmpegagent-native

Warlock Video Producer

A local video editing engine an agent drives: an edit plan becomes a structured timeline, the timeline becomes a deterministic render, and every render is checked and receipted. Nothing renders directly from prose.

Why it exists

Baby Warlock’s older video tooling could trim, render and check a single clip. It could not assemble clips, add transitions or motion, mix several audio tracks, or carry out an edit plan. The Warlock Video Producer fills that gap as its own package, with no dependency on Baby Warlock; Baby Warlock is one client of it. The roadmap’s key finding was that the missing piece was integration, not intelligence: the vision tooling already existed but nothing could reach it.

How it works

An agent writes an edit decision list. The engine turns it into a structured timeline of clips and effects, runs a deterministic render, checks the result, and writes a receipt. What it can do that the old tooling could not:

  • Multi-clip timelines with cuts, fades, crossfades and dips to black.
  • Keyframed zoom and pan, blur reveals and speed changes.
  • Multi-track audio with ducking and two-pass loudness normalisation.
  • Title and image overlays.
  • A restraint gate, enforced by the engine rather than by good intentions.
  • Subject tracking: the frame pans through every observation of the subject, not just the first and the last.

What is proven on real footage

  • Vision review. Real frames go to a local vision model, which reports what it sees against a fixed vocabulary. On one candidate the audio and scene signals liked, it correctly disagreed: the window was a store menu and black loading frames.
  • Model-authored edits. A second local model decides keep, drop or trim for each window. On real footage it dropped a static window and trimmed a foliage-blocked one to the exact second the vision review had marked the subject clearly visible, where the old rule would have thrown the window away. Trims outside the real window are rejected, not clamped, and every fallback says so.
  • Style profiles. Content and style are separate. Playthrough keeps chronological order with a three-second minimum; Short is hook-first, 60 seconds, vertical 1080 by 1920.
  • A regression reel built from one real gameplay capture covers 8 of 10 categories. The two empty ones stay empty rather than being filled with invented cases.

Honest limits

  • Not yet reachable by voice: it is not routed into Baby Warlock’s commands, and that needs its own authorisation.
  • Authoring is keep, drop and trim; it does not yet reorder windows for story.
  • No montage or trailer styles yet (they need beat detection and an arc).
  • No transcription inside the engine, and no B-roll insertion.
  • One real domain proven so far: 7 Days to Die gameplay.

Field reports

Filed against this project

All reports

No reports filed yet. The first one lands when there is something to prove.