Video production · Creative direction · Production systems
One long-form interview goes in. Vertical reels, thumbnails, quote graphics and a transcript come out, cut to the word, framed to the speaker, and checked before anything reaches a reviewer.
A single interview can owe a trailer, a batch of vertical reels, three thumbnails, quote cards and a transcript. Doing that by hand doesn’t scale. Doing it naively breaks in ways nobody catches until a client does.
Automated cutting fails in specific, repeatable ways. Clips start a syllable late and swallow the first word. Audio bleeds in from the sentence before. A vertical crop centered on the frame puts the speaker’s ear in the middle and their forehead out of shot. Captions drift out of sync with the words they belong to.
None of those are rendering problems. They’re verification problems. The render succeeds, the file plays, and the mistake ships.
The design principleA stage that can’t check
its own work isn’t finished.
The encoding was never the hard part. The hard part was defining what a correct cut is, then refusing to let a model decide it alone. Three rules did most of that work.
Some jobs read the video and propose moments. Others go the opposite way: the script is written first and the footage is searched to match it. Knowing which direction a request travels settles most of the ambiguity before any work starts.
A vision model can tell you a moment is worth using. Only the word-level transcript can tell you where the sentence actually begins. Every in and out a model proposes is re-derived against the transcript before a frame renders.
Crop choices render as before/after contact sheets sampled across every shot, so framing is looked at rather than argued about in numbers. Approved offsets are recorded, so a rebuild months later reproduces the same frame.
Where a brief and the audio disagree, the system does not silently pick one. Both versions ship, with a short written note explaining the difference, so the people responsible for the brand make the call with the facts in front of them.
A representative delivery: one long-form interview master, five vertical reels, every one landing inside the brief’s own stated length range.
Brief target 75–90s
Brief target 25–30s
Brief target 40–50s
Brief target 60–75s
Brief target 45–60s
Delivered inside the brief’s own ranges
What the system caught is the part worth reading. The master switches between camera singles mid-sentence, and the singles are framed so alike (same wall, same lighting) that standard scene detection read them as one continuous shot. Reaction shots came back reframed at the wrong speaker’s crop offset. Remapping the master frame-accurately fixed it, and established the rule the reframing subsystem now assumes: offsets are per shot, not per person. The same speaker measured +380 in one section and −50 in another.
Two smaller calls went the same way. One requested line was never actually spoken as a single continuous sentence, so both constructions shipped for the team to choose between. Another began a single frame after the programme cut away from the speaker, so that reel deliberately holds on the listener, and it reads as a reaction rather than a mistake.
Every departure from the cut list shipped with a short written note alongside the reel it affected, so reviewers could see what changed and why before approving.
Every clip is tested for a clipped first or last word and for audio bleeding in from the adjacent sentence. Four of the clipping subsystem’s seven scripts exist only to verify output.
Crop decisions render as side-by-side images sampled at the head, middle and tail of every shot, so framing that drifts as the speaker moves is caught before the render, not after.
Work returning from another tool is checked as numbers: geometry, frame count against duration, and speech-versus-music levels. A music bed sitting two decibels over the voice passes a viewing and fails a measurement.
Anyone can automate the render.
The value is in the checks that catch what the render can’t.
For the technically curious. A request names a goal, not a tool: a campaign router picks the domain, a domain router picks the mechanism, and only then does any work begin.
Clips, trailer, thumbnail, graphic, transcript or delivery, chosen before any tool is named.
A video-native model reads narrative arc, eyeline and shot boundaries from a timecode-burned proxy kept 1:1 with the master.
Word-level timestamps become the coordinate system. Every cut, caption and quote is addressed in words, not seconds.
Cut points snap to measured word edges with padding derived from the surrounding silence, then re-encode for frame accuracy.
Vertical crops track the subject shot by shot, because one global crop is wrong the moment the director cuts.
Typesetting, atmosphere and end cards composite in one pass, then a QA sweep proves the result against the intent.
Resolve a goal to a domain, then a domain to a tool. No router does any work itself.
Read source material and propose what is worth using.
Cut, assemble, reframe, typeset, grade, compose graphics and ship.
The deliverable was not a script.
It was a repeatable production capability.
Questions about this project? Talk to my bot