01
Never over the face.
Type lives in the empty space beside the speaker. If there is no empty space, make some (next pair).
Do
Don’t
The rules that hold for every brand we edit, house looks and client kits alike. The look changes per brand. The craft does not. Each rule plays a good example next to a mistake we really see. The examples are drawn in the Editorial look. Your brand’s colours and fonts come from its design system. Read it once, then run the two-minute check before every export. This page shows the rules with pictures. The full rules with every number, the patterns (how we show each kind of thing) and the checker script live in HARNESS/craft; this page is rebuilt from there.
Where things sit. The face is the most important thing on screen, and empty space is what makes a frame feel expensive.
Type lives in the empty space beside the speaker. If there is no empty space, make some (next pair).
The speaker looks to camera right. Shift them left for the whole shot and put the words on the right. The reframe changes on the cut, never mid-shot. Two titles in one shot share one position.
A definition, a list, a pro tip and a subtitle all at once is four graphics. Let one go out before the next comes in. If the picture already says it, there is no header.
Type has a budget: a text block uses at most 40% of the frame. Nothing important inside 10% of any edge, and keep key text above the bottom 12% where YouTube’s controls sit. Bigger type is not more important. Media is the opposite: it takes the space (see Showing media).
A side gradient is about half strength at the edge and gone by mid-frame, and only when the shot needs it. Never darken the speaker to make the type pop.
A screenshot is cut inside its own border. No sliver of a window edge, rounded corner, scrollbar or neighbouring card at the crop line. Check each edge at 2x before you render.
What goes on screen and how it is set. A title is the point of the line, not the sentence.
The accent is where the eye goes. Each brand has one accent move. Its design system says which. One method, one word, the word the speaker stresses.
If it takes longer to read than to hear, it fights the voice. Five words or fewer a line, two lines, three at most. Cut words; never shrink the type to fit.
No orphans, and never end a line on “a”, “the”, “to”, “of”, “with” or “and”. Centered lines make a shape: keep it a clean one.
Big type is tracked tight. Lines stack as one block. Contrast comes from size, not from piling on weights. One thing reads first.
Text shows on the word and holds long enough to read twice. Then it leaves with the shot, not after it.
One camera under the whole graphic. Text tied to a shot ends exactly on the cut, with Animate out off. Watch the cut mark at 2.4 s.
Calm. Eased, never linear. One entrance style per video. The step rate is the brand’s choice (its design system says which); the calm is everyone’s.
Each word rises into place and lands. Nothing slides in from off screen, spins or bounces.
A template sits where it is placed. No slide by default; if the text would touch the head, a small slow slide done before the first word. If a template can’t do it, that is a custom graphic, not a hand edit.
Studies, PDFs, screenshots, posters, diagrams, recordings and clips. Readable, sharp, moving, and the one thing he quotes lit. Rules M1 to M12 in HARNESS/craft.
What the viewer must read is at least 30 px cap height at 1080 at its moment. If the layout can’t get it there, use the full-screen pattern.
Never shown above its real pixels at delivery size. PDFs at 600 dpi or more (900 for deep zooms). Images at 2x the size they’re shown.
Freed space goes to the media: move it up, scale it up. Never parked off to one side with dead space around it.
Only the words he says, wiped on as he says them. No second stat or callout competing.
A wide sharp band, a long soft falloff, light blur and dim. Never a spotlight.
His face tile overlaps the media’s corner, fully inside the frame. Never a separate small window beside it.
Every still keeps moving: a slow 2 to 3% push, or beats on his words. The camera reads the media in the order he speaks.
Publication and year in a small chip. Real assets only: never a faked screenshot.
Every video gets one lens over the whole frame as the last step, so the type sits in the picture. Which lens is the brand’s. Rules X1 to X3.
An adjustment clip on the top track after lock. One lens per video, never stacked, nothing else adding grain or glow.
Compare none against the lens on a title frame, cropped at 100%.
Colour fringes on type. Glow that pushes skin red. Glow or grain in a light room.
Filmed, not pasted: soft shadow, slight softness, grain through the letters. Shade only as much as contrast needs, wide and soft, never a box, never on the face.
How we show each kind of thing, the same in every brand. Each has its own file in HARNESS/craft/patterns with layout, timing, camera and finish. The brand only dresses it.
P01 Side Title · P02 Name Card · P03 Definition · P04 Spoken List · P05 Captions · P06 Hook Captions · P07 Text Hook (shorts)
P10 Section Path (the video’s list, current part lit) · P11 Film Title · P12 Hook Teaser. No flat colour plates as chapter cards.
P20 Study Full Screen · P21 Study Window with face tile · P22 Poster Walk · P23 Stage Board · P24 Showcase Window · P25 Overhead Board · P26 Clip Panels · P27 Sponsor · P28 Montage Conveyor · P29 Shorts Canvas
P30 Push and Punch · P40 Finish Pass
Rough-cut craft that holds for every video, any brand. The full rulebook is HARNESS/scripts/v3/cut_rules.md.
Cut on phrase ends. They are also the best camera switch points. A picture change near a cut snaps onto the cut; otherwise it sits in the pause before the next word.
No shot under 24 frames. 20 for a reaction.
Cover with a second angle, a punch-in or a cutaway, in that order. Punch-ins only inside a chain of jumps.
An overhead or demo shot stays until the drawing or action is done.
A silent stretch over 2.5 s is trimmed to about 1 s, cut in the middle of the gap.
Flagged, never cut. The director decides.
Re-check every join. No clipped words. Each voice matches its own raw mic.
Every round starts from a copy. Never edit the director’s timeline; keep every hand edit they made and say what you saw.
Sparse, timed to the motion, at the director’s level. Each brand picks its own sounds by ear; this is how they are used.
At most one sound every 10 frames. It starts a few frames before the move and peaks as the move lands.
A soft whoosh when something rises in. A click or shutter when a card, label or photo lands. A ratchet ramp-up for slides and scrolls, its slow-down for counters. A glass pop for text or a glow appearing.
Match the director’s placed sounds, about -17 to -20 dBFS peak. Whooshes are the usual offender: pull them back to about -10 to -12 clip gain.
Beds about 18 dB under the voices (20 for shorts). Loud drum intros lower again. End on the song’s own stop or phrase end; never jump into a section with new instruments.
Two endings or two mixes: render both and let the director pick by ear.
Editorial’s list is in HARNESS graphics_rules G12. Signal and Reel get theirs after their first edited video. Until then, ask.
The project runs at the footage’s frame rate, not a fixed 23.976. A stepped look must divide that rate evenly: every 2 frames at 24 or 25. Render one title at a new rate before building graphics.
Rendered with a see-through background (ProRes 4444 with alpha), encoded on the Mac, on a track above the shot like a title. Keep the source so it can be re-timed.
Check phone footage for HDR at ingest. Brand colours come from the brand kit, never from memory.
A stat on screen must be provable. Never fake a screenshot. Only real quotes in a quote. Clips and thumbnails come from the client’s assets folder.
Before every export. If any line fails, fix it first.
One thing reads first, and it is the right thing.
Nothing touches the eyes or mouth. The text is on the side the speaker faces.
Type within its budget and the 10% margin. Media big and centred, nothing parked off to the side. Nothing over a head.
None, or wide and soft. Never a box. Never on the face.
One alignment edge, one margin, the whole video.
Tight tracking, close lines, two weights, one accent per line.
Lines break on meaning. No orphans. No “the” at a line end.
No graphic crosses a camera change. Text tied to a shot ends exactly on the cut. The checker script tests this before render.
Readable, sharp, highlighted only where he speaks, always moving, source named.
The brand’s lens is on as the last layer.
One entrance style. Nothing flies, nothing flashes, no keyframes on templates.
Every line it was given is on screen. No leftover example text. No unprovable stat.
One sound per move at most, timed to it, at the director’s level.
The project runs at the footage’s rate and the step divides evenly.
One contact sheet of stills, each graphic at its fully-in frame, looked at before anything is shared.
For every brand: HARNESS/craft (README = the map, rules.md = every rule with its numbers and test, patterns/ = one file per move, scripts/craft/craft_check.py = the gate). Changed there first, then this page and the Notion checklist, then one line in HARNESS/CHANGELOG.md.
Each brand’s colours, fonts, accent move, motion flavour and sounds: its design system. The link is in the Notion Template Packs table (Brand Kit [ Koi Media ] > Template Packs), Design system column.
Which template for which moment, and how to ask for it in a Frame.io comment: that template’s Notion page.
Stages, gates and how Claude works: the video-pipeline skill and the HARNESS rulebooks. Where a video stands: its STATE.md.