QuorX · Research · 20 October 2026

The Director's Harness

Agent-made video looks directed when film craft and the harness work together. The craft decides what the viewer feels. The harness plans the whole film before building it, renders the same frames every time, and critiques what it actually rendered.

Plan the whole film first

Someone signs off on the shot list before a frame is built. Chained prompts spread early mistakes.

Same frames every time

Each frame is a function of time, so the preview, the critique and the final render show the same film.

Critique the render

Vision critics review real frames, one topic each. Every candidate is judged against the original brief.

Analysis only. Everything here comes from reading public sources and the QuorX source code. Nothing was tested on video, and no testing has been done to prove how the decision models perform on video questions. Scores, estimates and projections are judgments; the decider numbers come from logs of mostly text benchmark tasks. Questions and corrections are welcome on PR #110, which adds the video skills.

Pre-productionSC 1

Shape the film before building it

Trailers and launch films share one shape: a cold open, an introduction, an escalation, a climax and a closing button. Below is a 30-second plan on that shape. Press play, click the timeline, or pick a shot from the list.

Fig 1.1 · Edit timelineExample plan · 30-second launch film

S5 00:00:11:06

Feature one, close up

Why it matters

Shot: S5 · 0:10.0 to 0:14.0
Act: Act 2 · Escalate
Purpose: Escalate
Framing: Close-up
Move: Push-in
Cut out: Cut on the bar line

Bar line every 2 s (120 BPM) Product on screen Sound effect

Acts follow the trailer structure. The music changes mood at three planned points (flags on A1), cuts land on bar lines, captions are burned in on their own track, and the product is on screen before 0:03.

Hook in three seconds. Show the product within 3 seconds and keep launch films to 15–30 seconds. EveryInc
One shape. Cold open, introduce, escalate, climax, button. For a product, that is the problem, the product, the payoff and the call to action. Oseman MasterClass
Cut to the music. Choose the track first and cut to it. It should change mood at only two or three planned points. Oseman PremiumBeat

Storyboard first, code second. Rewriting scenes nobody needed is the costliest mistake, so the shot list is reviewed before any scene is built. EveryInc
Real copy only. Use the product's real UI text. Invented labels are the top sign that a video feels fake. EveryInc
Hook, build, peak, resolve. Motion-design skills frame short videos the same way, and name over-animation as a common failure. motion-designer

Fig 1.2 · Shot listClick a row to cue the timeline

Shot	Time	Act	Purpose	Framing	Camera move	Cut out

Every shot states its purpose, framing, camera move and the cut out of it. The fields follow the storyboard format of the ai-video-storyboard-skill, which plans 6–18 shots per film.

ProductionSC 2

Keep the viewer oriented

Continuity rules matter whenever a film has real space: live footage, a 3-D scene with a virtual camera, or several generated shots of the same characters. Drag the red camera to see what each rule protects.

Fig 2.1 · Line of action, seen from aboveDrag the red camera or use the slider

Shot 2 camera position

Shot 1 · fixedA left · B right

CUT

Shot 2 · yoursA left · B right

180° rule. Keep the camera on one side of the line between the subjects. A cut across it can disorient the viewer, unless that confusion is the point. Adobe
30° rule. Move the camera at least 30° between two shots of the same subject. MIT film lexicon
Match on action. Cut in the middle of a movement. The viewer follows the action and misses the edit. MIT film lexicon
Decide sides in the shot list, especially for dialogue scenes. FilmSkills

ProductionSC 3

Ease every move, lead with one thing

Motion tells the eye where to look. The rules are small and specific: ease in and out, keep durations short, and stagger elements so one of them leads.

Fig 3.1 · Easing labCurves at slow motion · frame at real speed

ProgressTime

Ease-outentering · slows on arrival
Ease-inleaving · speeds away
Ease-in-outmoving within the frame
Linearlooks mechanical

Launch dayTitle card

Enter · ease-out · 300 ms

Entering elements slow down as they arrive, leaving elements speed up, and moves within the frame do both. Switch to Linear to see why constant speed looks mechanical. madegooddesigns

Fig 3.2 · How long a move should take

Most UI transitions sit between 200 and 500 ms and micro-interactions between 150 and 250 ms. Faster gets missed; slower feels sluggish. The red line follows the duration picked in Fig 3.1. madegooddesigns 75secondes

Fig 3.3 · One element leads

All at once

Staggered 110 ms

When five elements arrive together, nothing leads. Staggered, the eye follows the order you chose. Keep one main focus per frame, and offer gentler cross-fades for viewers who reduce motion. madegooddesigns animation-principles

Post-productionSC 4

Cut for emotion first

Walter Murch's Rule of Six ranks what a cut has to get right. When a cut can't satisfy all six, give up items from the bottom of the list first. If the emotion is right, viewers forgive the rest.

Fig 4.1 · Murch's Rule of Six, 100 squaresHover a criterion

1EmotionIs the cut true to what the audience should feel now?51%
2StoryDoes it move the story forward?23%
3RhythmDoes it land at the right moment?10%
4Eye-traceDoes it follow where the viewer is looking?7%
52-D planeDoes it keep the flat screen layout, such as screen direction?5%
63-D spaceDoes it respect the real space of the scene?4%

Keep from the top. Sacrifice from the bottom.

Weights as commonly cited from In the Blink of an Eye: emotion alone outweighs the other five. These percentages are repeated by course pages and blogs; they could not be checked against the book itself. UT Austin No Film School

Post-productionSC 5

Finish for a muted phone

Many viewers meet the film on a phone, sound off, under the platform's own buttons. Finish for that case first.

Fig 5.1 · Loudness targets by sourceLUFS, integrated

Most sources land on −14 LUFS with true peaks near −1 dBTP. For Instagram and TikTok they disagree: one vendor says louder, another says quieter. None of these are official platform specs. Opus video-use

Fig 5.2 · 9:16 safe zone

Show platform UI

1Top bar: tabs and search.
2Action rail: likes, comments, share.
3Bottom band: username, post caption, music label.
Keep text, faces and calls to action inside the dashed zone.

Zones are approximate; the sources give no exact pixel margins, so check each platform's current guide. Tella Reap

Fig 5.3 · Delivery spec

Item	Target	Why	Source
Text	Readable at 720p	Feeds and previews downscale the master	EveryInc
Captions	Burned in	Many viewers watch with the sound off	EveryInc
Landscape master	1920×1080 · 30 fps · H.264 CRF 18	A clean master for every later cut	EveryInc
Vertical cut	1080×1920 · 9:16	Native vertical gets the most reach on short-form feeds	InfluenceFlow
Loudness	−14 LUFS · −1 dBTP	Louder masters get turned down by the platform	Opus video-use
Copy	The product's real UI text	Invented labels make a video feel fake	EveryInc

The harnessSC 6

The loop that makes it look directed

The agent systems that turn out well-directed video share one loop. They plan the whole film, lock the style, build from tested recipes, render deterministically, critique the real frames in parallel, and pick the winner against the original brief.

Fig 6.1 · The common loopClick any step

Step 1 of 15

Brief + real assets

Assembled from EveryInc's launch-video skill, OpenMontage, video-shotcraft, HyperFrames, Google's co-director and VISTA. The amber dashed line carries the original brief to the final selection.

Fig 6.2 · Why rendering must seek

The same ten frames with the same uneven render times (grey bars). Animation that reads the wall clock jumps at frame 4 and finishes early, so frames 7–9 repeat. Seeking frame n puts every ball where it belongs, however long the frame took. HyperFrames Remotion

Fig 6.3 · Skills load on demand

HyperFrames' router maps a request to one of its creation workflows, and the workflow loads domain skills only when it needs them. Remotion's /remotion-best-practices plays the same routing role. Which domain skills light up is illustrative. HyperFrames remotion-dev/skills

Fig 6.4 · Harness checklist0 of 12 ticked

Practice	Why it matters	Seen in
Plan the whole film, then build	Chained prompts spread early errors and make failures hard to trace.	EveryInc OpenMontage Google
Sign-off before costly steps	The person can redirect before time and money are spent.	OpenMontage EveryInc
Deterministic, seekable rendering	Preview, critique and final render all see the same frames.	HyperFrames Remotion
Router skill, domain skills on demand	Only the rules a task needs sit in context.	HyperFrames Remotion
Shot-recipe library	Tested timing, easing and choreography instead of improvised moves.	video-shotcraft
Lock style and references early	Characters, palette and type stay the same across shots.	Google invideo
Lint and check inside the loop	Catches animation that can't seek and broken files before review.	HyperFrames video-shotcraft
Vision critique with a rubric and a hard limit	Catches what text-only checks miss, and the loop always ends.	MindStudio
Critics split by topic, on another model	Each critic has one job; a different model avoids shared blind spots.	EveryInc VISTA Taskade
Select against the original brief	Stops local fixes from drifting away from the intent.	Google VISTA
Editable output as tracks	People can adjust shots, captions and sound afterwards.	video-shotcraft
Confirm before posting, dry run first	Posting is public and hard to undo.	video-shotcraft

Ticks are saved in this browser only.

ResearchSC 7

What the research systems add

Two 2026 systems from Google Research push the loop further. One searches over creative directions; the other runs a tournament between candidates and critiques the winner.

Fig 7.1 · Google's AI video co-directorTwo loops

The inner loop turns one creative configuration into a storyboard, keyframes, clips and audio. The outer loop feeds the judge's score back to a multi-armed bandit that picks the next configuration. CANVAS memory keeps characters and places stable when scenes return. A final global selection re-scores every candidate against the original prompt. Google Research

Fig 7.2 · VISTA's tournament loopCycle 1 · planning

VISTA plans scenes, generates candidates, keeps the winner of a pairwise tournament, critiques it for visual, audio and context fidelity, and rewrites the prompt for the next cycle. This diagram follows the abstract; the full paper could not be opened. VISTA

Fig 7.3 · VISTA's reported results

From the abstract. The line marks 50%, a coin flip.

MVAgent passes typed conditions between agents instead of repeating text, and keeps a continuity memory of how each shot ends. arXiv

MovieAgent plans a movie with multi-agent chain-of-thought before generating. GitHub

The lessons Google draws: treat consistency as its own objective, plan globally, keep explicit world-state memory, critique in words, select against the original intent, and keep safety layers such as watermarking in the stack.

ToolkitsSC 8

What you can install today

Most toolkits are community projects for coding agents. They differ in what they render with and how much of the loop they cover.

Fig 8.1 · Toolkits compared6 of 6 match

Toolkit	Renders with	Plans first	Review step	Library	Sound	Worth knowing
Remotion skillsofficial	Remotion (React)	Not stated	Not stated	12 skills; /remotion-best-practices routes to the rest	Audio and captions topics	npx skills add remotion-dev/skills
HyperFramesHeyGen	HTML in headless Chrome + FFmpeg	Beat planning skill, no gate	lint, check, snapshot, keyframe diagnostics	21 skills behind a router, plus a block registry	Mixing with voice-over ducking	Apache 2.0; seeks every frame; no build step
video-shotcraftcommunity	Remotion	Storyboard step	Frame checks with ffprobe, aesthetic rules	124 shot recipe cards in 10 categories	Beat sync, 149 SFX	Workbench edits shots, captions and SFX as tracks
product-launch-videoEveryInc	Remotion scaffold	Shot-list table first	4 parallel critics: design, readability, pacing, brand	Starter scaffold	Not stated	15–30 s; product by 0:03; 1080p30, CRF 18
OpenMontagecommunity	Remotion, Python tools, Piper TTS	Treatment before costly steps	Not stated	12 pipelines, YAML manifests, hundreds of skill files	Text-to-speech voice	AGPL-3.0; counts vary between write-ups
ai-video-storyboard-skillcommunity	Doesn't render	6–18 shot plan	Post-production checklist	None	None	Writes prompts for AI video models

Read from each project's README or docs in October 2026. "Not stated" means the page didn't say. Check each repo before installing.

Fig 8.2 · 318 repos by type

One curated list of video skills for coding agents, by type. Only 10 repos focus on pre-production: scripts and shot analysis. curated list

Fig 8.3 · One brief, four render times

9 of 16
skills tested added sound

From the same list's own test: one brief, one run each. Treat it as a single data point, not a benchmark.

ScorecardSC 9

How QuorX measures up

QuorX is a general build harness, not a video tool. Its loop for apps, sites and games already matches or beats most of these practices. What it lacks is everything specific to video: a frame renderer, a timeline, video craft, and checks on rendered frames.

Fig 9.1 · Scorecard against 13 practicesFrom the QuorX source, October 2026

Status of the 13 practices

Practice	Best practice	QuorX today	Status	Distance

Points: ahead or matches = 2, partial = 1, missing = 0. "Ahead" means QuorX enforces or measures the practice beyond what the toolkit READMEs describe. Evidence comes from reading the QuorX source; none of it was benchmarked for video.

Fig 9.2 · Where QuorX is already ahead

An enforced plan gate

Build tools are refused until the person says Go ahead to the plan. The toolkits only ask the agent to storyboard first.

server/agent.js

A reviewer that is measured itself

Every verdict is logged and compared with a judge, and the default reviewer comes from a different model family than most builders.

logs/reviews.jsonl · scripts/checks.mjs

Hardened approvals

The judge sees only the person's own words and facts the harness computed, never the agent's text. It is never the chat's own model, and split votes go to the person.

server/approvals.js

A prompt enhancer

Spells out what a request implies ("a 30-second movie" means voices, music and sound) for the agent and the judges to read.

server/enhancer.js

Fixed harness checks

Decider questions at set moments: before a build, after a screenshot, after a click. The harness writes them, not the building model.

server/harness-checks.js

Priced media behind approval

Paid images, video clips, music and speech each need approval, and the cards show prices measured on earlier runs.

server/agent.js · server/budget.js

Fig 9.3 · Closing the gapsEffort and order are estimates

Nowsmall changes

A video-direction skill

Story shape, shot lists, 180° and 30° rules, easing, the Rule of Six and the delivery spec, written from this page.

Closes: video craft

Topic critics

Run the reviewer as three or four parallel critics, each with its own rubric, on decider models.

Closes: split critics

Checks on rendered files

Add ffprobe and a seek test to the write checks, so a broken render fails before review.

Closes: lint in the loop

Nextmedium changes

A shot-recipe library

Seed it with the trailer's own shot list, one card and one runnable demo per shot.

Closes: shot library

Candidates judged against the brief

Render two or three cuts and pick one by tournament, judged against the enhancer's reading of the person's words.

Closes: select vs. brief

Laterlarge changes

A render tool

Seek each frame in the headless browser and encode with ffmpeg as a background job, so QuorX can make code-drawn films itself.

Closes: deterministic render

Tracks the person can edit

A timeline manifest and editor for shots, captions and sound effects.

Closes: editable tracks

Projected score after each stage, using the same points as Fig 9.1. The loop is already strong, so most of the climb comes from the video stack.

DecidersSC 10

Slotting the deciders into the video loop

QuorX's deciders are fast decision models: typesafe/jev-1.14, liquid/d1-ultra and openai/gpt-6-luna-decisions by default, up to five. They don't write notes. They answer typed questions with probabilities in about half a second each, and QuorX averages their votes. That suits the dozens of small yes-or-no calls a video loop needs, about text or about frames, and leaves the writing of fixes to the builder and the reviewer.

Fig 10.1 · One question, three decidersExample votes · real formulas and latencies

noul · yes/noimage

Builder's guess, committed first

Average3 of 3 votes

Confidence0.00

0.5 lean0.8 settled

◐Uncertain

Each decider gets every question in one request; choice options are shuffled per decider, because some lean toward the first option. Confidence uses the per-type formulas in server/decisions.js: |2P−1| for yes/no, the top share above chance for a choice, and the spread around the top level for a score. Settled at 0.8 or more, lean at 0.5. The bars arrive at each decider's median response time from the maintainer's QuorX run logs.

Fig 10.2 · Where deciders fit in the loopText and image slots

in QuorXalready asked todaynew slotwould close a gap from SC 9image judged from framestext judged from words

Brief
in QuorXtextyes/no
Gap check, with video fields
"The message says how long the film is." "…where it will be posted." "…whether it needs a voice." Today's gap check asks what a build leaves out; the prompt enhancer already spells out that a movie implies voices, music and sound.
Settled no → ask the person one question card before planning.
Shot list
new slottextyes/no
Plan sanity before sign-off
"Every shot states a purpose and the cut out of it." "The product is on screen before 0:03." "Music changes mood at no more than three points."
Any settled no → the builder fixes the plan before the person sees it.
Style lock
in QuorXimagechoice
Read a reference's look
extract_design already has the vision deciders judge a picture's background, accent, type style, corners and density. The same call can read a mood board or a reference film still.
Answers become DESIGN.md and tokens.css, which every later check reads.
Frames
new slotimageyes/no
Frame checks on the contact sheet
"The caption in this frame is fully readable at phone size." "Nothing important sits under the right-hand rail." "A is on the left in both frames." One look decider answers (gpt-6-luna-decisions by default) and the others join only when it is unsure.
A settled no names the frame and the rule, and goes to the builder as a bug.
Critique
new slotimagescore
Topic critics as score questions
One score question per topic (design, readability, pacing, brand), all deciders in parallel on the same sheet. The reviewer model writes fix notes only for the topics that score low.
Fills "critics split by topic" for about a tenth of a cent a round.
Select
new slotimagechoice
Pairwise tournament against the brief
"Which of these two sheets better fits what the person asked: '…'?" with the person's own words and the enhancer's reading, never the builder's notes. Shuffled options cancel first-position bias.
Three cuts take two matches; the winner goes to the final render.
Spend + post
in QuorXtextyes/no
Approval judge for paid media and posting
make_video, compose_music and speak already go through it: a quorum of deciders that never includes the chat's own model, seeing only the person's words and the harness's facts.
Split or unsure → the person decides.

Leave to measurement what can be measured: loudness, beat sync, durations and frame timing come from ffmpeg and ffprobe, not from a vote.

Fig 10.3 · Speed per decidermedian ms per vote

From 531 decider votes in the maintainer's QuorX run logs. A question goes to all deciders at once, so it takes as long as the slowest: about 0.9 s with the defaults.

Fig 10.4 · Cost per decider$ per 1,000 votes

A three-decider question costs about $0.0003. A 12-frame contact sheet with three checks per frame is about $0.005 on the look decider alone.

Fig 10.5 · What 868 logged questions showlogs/decisions.jsonl · mostly benchmark text tasks

Verdict

Deciders agreeing with the pick

Against the builder's committed guess, 72% of questions were ones it already knew: asking was theatre. The 49 changed answers are where the deciders earned their cost. For video that means asking about pixels the builder can't judge well (legibility, safe zones, screen direction), not facts it wrote itself.

Guess first, then ask. The builder commits an answer and a confidence, so the log shows which questions changed anything. Fig 10.5
Keep the builder's claims out. In QuorX's benchmark, a model's confident misreading pulled the deciders to P=0.77; labelling its notes unverified, with "the user's message wins", brought the same deciders to P=0.04.
Shuffle every choice. Already done per decider, and it matters most in a tournament, where position bias would pick the first cut.

Don't let a majority of drafts overrule a confident pick. Drafts from one model share their mistakes: the majority was right 1 time in 5 against a confident pick, which was right 3 times in 3.
Quorum for money and posting. One model's parse must never stand for the panel; the approval judge already requires a quorum of valid votes.
Measure, don't vote. Loudness, beat sync and durations have exact answers; give the deciders only the judgments a person would make by looking.

Update · 23 OctSC 11

More than one way to make a video

A video can be drawn in code, built from still images, made from AI-generated clips, cut from real footage, or a mix. So the single video-direction skill became ten: a hub that picks the route, one skill per route, and shared craft for style, sound and review. They're proposed in PR #110 and still untested on video.

Skill	Route or craft	Covers
video-direction	Hub	Picks the route; brief, treatment, story, camera and motion rules; delivery
film-render	Code	A page that draws any moment, frame capture in a headless browser, ffmpeg
film-from-images	Images	Photos or generated stills with camera moves and parallax
film-from-ai-clips	AI clips	A spend plan, one look line, start images for a consistent cast, choosing takes
film-from-footage	Footage	A transcript-led edit, b-roll, word-timed captions, a timeline your editor can open
motion-graphics	Code	5–15 s design-led pieces with checkable signature features
three-from-image	Code, 3-D	Props and characters in three.js code, compared with the reference after every pass
film-style	Craft	Reusable looks, proven with a style frame and a short demo
film-sound	Craft	Ambience, foley, an original score, voice checked by transcribing it back, a −14 LUFS master
film-review	Craft	Measured checks, look checks, an anchored 1–10 rubric, fix rounds

Where the ideas came from. Mainly lemo-opuscar and mg-styles-15 (both MIT) and img2threejs (Apache-2.0).
Still analysis only. None of the skills has been run on a real video task in QuorX yet. The research discussion in issue #33 is closed now that its findings are written up.

End creditsSources

Skills and toolkits

Research

Craft

Delivery

QuorX newsletter

Get the next piece of research.

One email when there's a new QuorX release or research like this. Kept on our own server, never sold, and you can leave anytime.