The Order Is the Rule

The whole method fits in 28,498 bytes, 4,443 words and 402 lines, and it opens with 1 chain of 14 stages: brief, audience, 5-line story, script, shot design, moodboard, stills, approval gate, voice, music, video, clip gate, edit-room prep, deliver. Under the chain sit 16 numbered phases, 19 sub-headings, 13 tables and 10 code blocks. Every phase names 3 things: the skill that governs it, the artifact it has to produce and the gate that has to pass before the next phase opens. The load-order table alone is 12 rows and names 17 distinct skills. The cost of skipping a gate is written in 1 line, and it is arithmetic rather than opinion: a 6 out of 10 still that still gets animated burns more money than a re-render. Read the document end to end and almost nothing in it is a style preference. Phase 1 is measurement. Phase 6 is naming. Phase 8 is a gate. Phase 13 is 27 crossed-out lines. The 402 lines are mostly checks against something written further up the page, which is the only reason the same film can be produced twice by two different people.

Phase 1 is 1 instruction: pull the live site and measure it. Not the copy, the stylesheet. A homepage can be a JavaScript shell, so the real identity of a brand lives in the hex tokens, the font declarations and the logo file. The phase produces a 10 row constant table: background, accent, text, muted text, display font, mono font, brand line, product surfaces, real proof and register. From that moment the table is a hard constant and every prompt downstream inherits it. Two rules sit beside it. The model writing the prompts is text-only, so image reads are routed to a vision model instead of being assumed. And direction gets locked with the client before a single generation credit is spent: which format, which world, which facts are public and which stay blurred. The failure cost is stated plainly in the same section. A guess at phase 1 turns into a re-render at phase 9, and a re-render at phase 9 costs more than the measurement ever would have. 10 rows, 1 hour of work, and it is the entire difference between 1 film and 3 attempts at the same film.

Story is settled in 5 lines before any prose exists. Situation names 1 operator, 1 place and 1 belief. Desire is a verb. Conflict is a blocker that could actually refuse. Change is 1 moment. Result is the same place in a different state. Then 5 tests run and each one is pass or fail: does the situation carry character, place and belief in 1 breath, is the desire a verb, can the conflict refuse, is the change 1 panel, and is the result genuinely different from the situation. Every line of script then has to trace back to a card. No card, no line. The audience table above it holds 4 rows and only 1 of them is allowed to win, because the arc flips depending on which row is chosen: the risk-averse buyer wants verification, the funder wants provenance, the passing visitor wants 1 idea and no jargon wall, the recruit wants proof the work is real and not theater. After that the body is 6 beats ordered by rising impact. Never open on the strongest beat. The strongest beat is the payoff, and a payoff in position 1 is worth nothing.

Every shot has a name before it is designed. The library holds 21 named shots and 7 memory anchors, and the worked example used 7 of them: 1 anchor still that locks the world, then 6 story shots built image-to-image from it. The anchor is not a nice-to-have. Building every still from 1 reference is cast lock by construction, and it removes the drift that otherwise shows up 4 shots later when nobody is checking. Then 7 framing rules run as hard constraints: 1 named light source with every competing source killed, hands described by start and end position and never by the path between them, held faces capped at 3 seconds, negatives carrying no text and no watermark, framing in the 35 to 50 millimetre range, and 1 accent element per shot with the accent appearing only where the system verifies. Every still ships before 1 second of video is generated. Retries are capped at 4 per still and rejects get archived rather than deleted. The gate on each still is 2 checks and 5 dimensions. Deterministic first: a split-mirror correlation reading above 0.72 for duplicates and above 0.78 for mirrors, plus a mean-luminance light check. Then vision: consistency, colour, storytelling, integrity and impact. Below 7 out of 10 on impact is a reject unless it is fixable.

One document: 28,498 bytes, 4,443 words, 402 lines.

The failure register is 27 crossed-out lines in 4 groups, and it is the most valuable page in the file because none of it is theory. Story and script: no intro written before the outline, no strongest beat first, no call to action parked at the outro when the viewer is already leaving, no second call to action, no banned-word list waved through. Visuals and rendering: never generate video from an ungated still, never interleave the 2 phases, never mix warm and cool light in one frame because it drifts, never re-describe the cast in every prompt, never hold an apex frame for 6 seconds when it decays, never ship a fake dashboard or an invented acronym on screen. Audio: music never swells over voice, 1 continuous take never lands on a montage, an instrumental bed gets forced instrumental or a singer arrives on it, and the bed sits between minus 18 and minus 24 LUFS under the voice with the master at minus 14 and a true peak no higher than minus 1. Edit room: no blind placement through the append call, no trimming after placement, no rebuilding audio when a mixed master already exists, and no claiming a render finished because the queue said so. 27 lines, every one of them paid for somewhere.

Here is the arithmetic of the whole thing. The document is 28,498 bytes. It holds 14 stages, 16 phases, 13 tables, 71 bullet lines and 8 final QA gates. It carries a 27 line failure register and a 4 row audience table. Every one of those numbers exists so that a decision is taken once, upstream, in writing, by someone who is still allowed to change it. The delivery spec is 4 lines: 1920 by 1080 or 1080 by 1920, 24 frames per second, H.264 with AAC at 192 kilobits per second, and 3 checkpoint frames inspected per clip before anything enters the edit room. Read the 8 gates in phase 14 and notice that 6 of them are reading rather than rendering: read the copy aloud and see whether a stranger can repeat the essence and 1 number, check every on-screen number against the public record, quote the captions back exactly, watch the master once as the target buyer would. Entrade built 200 wood gasification machines on the same principle, which is that the second unit has to cost less than the first and the only way that happens is if the order gets fixed before the work starts. The probability that a pipeline with no gates ships a cheap frame into an expensive phase is not small, and the bill arrives as re-renders. That is why the order is the rule. Keep more of yourself, for longer.