One Order For The Whole Pipeline

Every stage of a social video is decided before the shoot. The hook, the cut rate, the framing, the voice, the grade, and the file that ships. 6 stages, and the order that governs them is written down before the first frame exists. Loaded late, those same 6 stages become 6 repairs. The hook is re-shot. The cut rate is fixed in the edit. The framing is saved by a crop. The voice is re-recorded. The grade is pushed to hide what the render got wrong. The file is exported at a size nobody chose. Loaded early, it is 1 document and about 20 minutes of reading. The playbook is 1,563 words. It covers the algorithm, the hook stack, the cut grammar, the voice stack, the visual limits of the model, and the render pipeline. It sits at the top of the job, and every later decision answers to it. The order matters more than the talent inside it, because a pipeline is only as strong as its weakest stage. A perfect hook on a video that dies at 3 seconds because the premise was unclear is a lost video. A clear premise under a frame nobody can look at is also lost. So the first move on any production is not a render. It is a read. Load the order, then answer 6 questions in writing. Who is this for. What does the first frame show. What is promised by 3 seconds. How often does the cut land. What voice carries it. What file is being verified at the end. 6 questions, 1 page, before any model is touched.

Views are no longer the unit that decides the outcome. Behaviour inside the view is, measured at 4 points, and each 1 is tracked separately. At 1 second the only question is whether the frame was worth staying for. The visual registers before the audio, so the first frame does work no sentence can do. At 3 seconds the premise has to be committed to. 63 per cent of the top click through videos hook inside 3 seconds. The average mobile decision takes about 1.7 seconds. Both numbers describe the same window, and neither can be bought back later. At the midpoint, delivery has to match the promise the hook made. This is where a strong opening and a thin body are found out. At the end, the payoff decides whether the video was finished, rewatched or shared. Intro retention above 70 per cent reads as strong. Below 40 per cent reads as dead. The quiet failure sits between the 2 bars: a video that dies at about 300 views did not lose an argument about quality. Nobody stayed past the hook. 4 checkpoints, each on its own, and any 1 of them can end the run. That is the reason the order is loaded first. A 1 second problem cannot be diagnosed at the export stage, and a 3 second problem cannot be fixed in the grade. The probability that a video recovers from a weak first frame is close to 0. The probability that a video with a clear hook, a matched midpoint and a real payoff reaches the next viewer is high. Both cost the same to shoot.

The hook has 2 jobs, and most videos only do the first. It stops the scroll, and it signals the niche so the algorithm routes the video to the right viewers. A confusing hook collects views from the wrong audience, retention falls, and the video dies with a healthy view count in its first hour. 3 layers, in order, inside 3 seconds. The visual trigger runs from 0.0 to 0.5 of a second. Movement, contrast, bold text on screen, a face with energy. Most viewers watch muted first, so the frame carries the load before a single word is understood. The verbal trigger runs from 0.5 to 1.5 seconds. 1 clear sentence: a promise, a curiosity gap or a problem. No greeting, no throat clearing. The context lock runs from 1.5 to 3.0 seconds. Who this is for, and what they get. That is the layer that tells the algorithm where the video belongs. 7 frameworks do the work: shock to logic, specific result, open loop, hot take, point of view, speed run, proof first. Each 1 is a shape rather than a line. The killers are written down too. A greeting intro, a generic promise, a slow setup, and no on screen text. 4 ways to lose a video before the content starts. Written down, the first 3 seconds are a checklist. Improvised, they are a guess taken under time pressure, which is the same thing as a loss.

The edit is where the order becomes audible, and 1 rule sits underneath it. The cut lands every 0.5 to 2 seconds, and most land at about 1. That is not a preference. Music at 110 to 120 beats per minute puts a beat roughly every half second. A cut that lands on the beat reads as intention. A cut that lands beside it reads as an accident. The same logic made the speed ramp on a snap zoom the most copied technique on the platform: the move has to have an end. Captions are effectively mandatory. Burned in, at 36 to 60 point or larger, centred or in the upper third, with the keyword marked. A large share of the audience watches with the sound off, and an unreadable video is a silent one. 2 older rules still hold under all of it. A cut between 2 shots of the same subject moves at least 30 degrees and changes size, or it reads as a jump. Stay on 1 side of the axis, or the screen direction flips and the viewer loses the room. Then the grammar that makes meaning. Cut on action to stay invisible. Cut on contrast to say something. 1 does the housekeeping. The other does the argument. Every 1 of those numbers is a decision made before the edit starts, and not 1 of them can be added by a filter afterwards.

The order is loaded before the first frame, not after the edit.

The largest single source of rework in the pipeline is a shot the model cannot hold. There are 4 danger zones, ordered by how badly each 1 breaks. Hands first. Extra and fused fingers are the number one tell. The fix lives in the shot list and never in post. Keep a hand out of tight focus. Describe a start position and an end position rather than a gesture. Slow the movement down. Use the back of the hand or a wrist crop when a hand is wanted in the frame at all. No fingers visible in the frame is the most useful sentence in the whole order. Faces second. An extreme close up pulls waxy and cratered skin, so a face shot sits at 85mm and medium framed. A static hold drifts past about 3 seconds, so it stays short or it moves. A head turn past 90 degrees smears. 1 hero face per frame, and 1 emotion at a time. Skin third. The beauty filter is the default, and the words that trigger it are the flattering ones: beautiful, flawless, perfect, radiant. The repair is the opposite list of 9 defaults, from documentary film and 35mm to natural skin texture, visible pores, unretouched, natural light, negative fill, slow deliberate movement and 1 subject. Add 7 biological cues, from pores to peach fuzz catching the rim light, and the render moves back toward a photograph. Text and reflections last. Fine patterns moire, mirrors and glass mismatch, and lettering garbles. Beneath all 4 zones sits the frame decision. 85mm medium for a face, 32 to 35mm over the shoulder, 35 to 40mm walking, 24mm wide and deep. Name the movement and its speed on every 1, or the model hands back a generic drift.

6 steps run from the first still to the shipped file, and the order covers all 6 of them. Stills first, always. Lock the identity, the body and the lens as a still, and gate it before anything moves. Never trigger video on a still that has already drifted. Gate before animate. The still clears the benchmarks or it is re-rendered. A broken still becomes a broken video every time, and no amount of animation hides it. Routing next, 1 model per job. Stills at 720 by 1,280, video at the size the plan allows, local renders for the hero shots. Assembly after that. Trim each clip to its window, apply the per cut move, join the pieces, burn the captions, mix the voice and the music, export 720 by 1,280 at 24 frames per second. Grade with 1 palette across the whole film, so the piece does not drift from shot to shot. Verify on disk last. The file exists, it probes clean, the dimensions and the frame rate are right, and it plays. An honest failure beats a claimed finish, because a claimed finish ships as a broken file. The voice sits inside the same order. A natural read or no read, and a robotic delivery never ships. 30 years of building, in 12 countries, under 2 rules: work with what the constraint allows, and refuse the version that looks better on paper than it can look in the room. Built companies across 12 countries. Restructured an industrial group. Deployed 210 energy systems. Took a company through a Frankfurt IPO. Built 200 wood gasification machines at Entrade. Signed a personal guarantee on a company that folded on Christmas Day 2008. In the building for the 2003 blackout. 393 pages in paperback, published 8 April 2026. Not 1 of those was carried by finding the rules late. The hook, the cut rate, the framing, the voice and the verified file are 6 stages of 1 decision, taken early. The probability that a pipeline which loads the order first reaches a shipped file is high. The probability that a pipeline which learns its rules in the edit does is close to 0, and it spends 4 passes finding out.