Every Cut Has A Named Reason

The failure this guide was written to fix is narrow and specific. Consecutive multi-shot renders cut like slides. Every frame is acceptable on its own and the sequence jumps, because a jump is not a style choice, it is a cut that broke a rule. The test is 1 line long. If a cut makes the audience notice the cut, a rule was broken. The cuts that read as invisible are not magic, they are 12 rules satisfied at the same time, plus 1 clause of reason attached to every single cut. The guide is marked decision-grade and apply before every render, which is the point of it. It is not a mood board, it is a checklist with a fixed order, and it is short enough to run top to bottom inside the 30 seconds before a render pair goes out. Under the test sit 13 numbered rules, 10 checklist lines and 1 meta-rule that governs the other 12. The material was pulled from continuity editing sources covering the 180 degree rule, the 30 degree rule and screen direction writing, plus 1 peer reviewed study on shot continuity in edited sequences and 3 multi-shot prompting guides from 3 different video models. The same grammar survived being translated from a physical set to a text prompt, which is why it still works when nobody is standing behind a camera. 2 counts are worth holding from the start. 13 rules. 1 clause per cut.

R1 is the 30 degree rule, and it catches more renders than any other line in the guide. Any cut between 2 shots of the same subject moves the camera 30 degrees or more in angle AND changes shot size or focal length. Under 30 degrees it is a jump cut by definition, whatever the frame happens to contain. There are 2 legal repairs when a room will not allow 30 degrees of movement. Change size dramatically, from the extreme close up to the wide. Or insert a shot of a different subject. Both close the problem, and neither requires moving the camera. R2 is the 180 degree rule, also called the axis of action. Pick the imaginary line between the 2 subjects and keep every camera on 1 side of it. Cross the line and screen direction flips, which disorients the viewer in 1 frame. The practical form is 1 habit: every shot prompt states which side of the axis the camera sits on. R3 is screen direction. A subject moving screen left keeps moving screen left across every cut, until an on screen turn or a re establishing wide resets it. Consecutive prompts therefore carry identical direction tokens, written the same way each time rather than paraphrased into something the renderer is free to reinterpret. 3 rules, 1 line of test each, and 1 shared property. All 3 are declared in the prompt before the render rather than repaired in the edit after it.

R4 is the eyeline match. Where a subject looks, in frame position and in height, has to match where the other subject sits in the reverse shot. If A looks down and left, the reverse places B up and right and shoots at a matching height. It gets stated in every prompt, because an eyeline that misses reads as 2 people standing in 2 separate rooms. R5 is the cut on action. Cut in the middle of a movement, never before it starts and never after it ends. In a generated sequence the form is exact: end shot N mid-motion, then start shot N+1 with the same motion already in progress, at the same speed, in the same direction. The movement is the bridge. R6 is the match cut. Shot N+1 opens on a shape, a motion or a composition similar to the final frame of shot N. The spinning top that closes 1 film and opens the next. The cut from a plane to the sea. It is reached for when the change is deliberate, because the rhyme makes the join read as designed instead of broken. R7 covers the 2 audio bridges. An L cut brings the next shot's picture in while the previous shot's audio is still playing. A J cut starts the next shot's audio before its picture lands. 1 overlap, 1 seam covered, and the sound leads the eye through the cut rather than arriving after it.

R8 is the repair kit, and it saves sequences that are already in trouble. When 2 shots of the same subject have to sit next to each other under a 30 degree change, something goes between them. A cutaway or an insert of a different subject. A match cut. A re establishing wide. What never happens is 2 same-subject, same-angle, same-size shots placed back to back. R9 is the size ladder. Adjacent shots differ by 1 perceptible step, wide shot to medium shot to close up. Dialogue reverse shots legitimately pair at a matching size, and that pairing is the 1 exception. 2 identical sizes at the same angle is a jump. R10 is camera motion continuity. If shot N's camera dollies right, shot N+1 continues that move, goes static, or justifies a new move. Reversing the direction of camera motion across a cut is the 1 thing that never reads as intentional. R11 is light and colour temperature continuity, and it is the most purely mechanical rule on the list. Consecutive shots share key direction, colour temperature and intensity, stated byte for byte in every prompt. Warm 3200K key from camera right, cool 5600K fill from left. Not close. Not similar. The same characters, 8 prompts in a row. R12 is temporal and object continuity. Time does not jump inside a scene. Hand positions, object states, wardrobe and hair carry over exactly, which means the prompt language is literal rather than evocative: glass half full, cigarette in the left hand, sleeve pushed to the elbow. 4 details, and any 1 of them drifting restarts the scene inside the viewer's head.

If a cut makes the audience notice the cut, a rule was broken. The entire discipline is 1 test, applied 1 cut at a time.

R13 is the meta-rule, and it polices the other 12. Every cut advances the action, changes information, or lands on a new idea, and the audience always knows where everyone is. If the reason cannot be named in 1 clause, the cut does not happen. The header that makes all of this reproducible is 9 lines long, and it is copied verbatim into all 8 prompts of a sequence. Character. Location. Light and colour temperature. Axis. Screen direction. Shot. Eyeline. Continuity state. Action. The renderer is only ever allowed to vary 2 of the 9, and those 2 are shot and action. Then 1 check runs before every render pair and it takes roughly 30 seconds. 5 questions. Is the same screen direction carried. Do the eyelines cross match. Is the size step at least 1 perceptible level. Are the light and colour temperature tokens identical. Is the motion continuous, with no reversed camera or subject movement. 5 questions, 9 slots, 1 clause. The discipline is smaller than the render queue it protects, which is the only reason a team will actually run it.

The applied proof is a montage of 8 shots, 2 minutes long, roughly 15 seconds each, worked shot by shot against all 13 rules instead of treated as 1 whole piece. The motif that holds the cuts together is 1 smile, moving from seen to forming to learning to complete. The global tokens are byte identical across all 8 shots. A warm 3200K key from camera right. Screen direction left to right throughout, with 1 exception. 35mm shallow depth of field on every character shot and 24mm on the 2 wide establishing shots. The protagonist appears in 4 of the 8, and the same 5 wardrobe tokens are repeated verbatim every time she does. 3 rules are broken on purpose, and each 1 is named in the same table that breaks it. The size ladder fails once, at the cut from shot 4 into shot 5, close up into close up, and the bridge is a match on the eyes plus a named colour temperature shift that marks a new era. The colour temperature changes once, at shot 5, when the warm key gives way to cool 5600K window light with warm 3200K city lights sliding past. The direction resets once, at shot 6, when the subject turns to camera for the only time in 8 shots. 1 transition sits outside the cut rules entirely. The closing shot holds 3 seconds after the smile ends, then fades. Every other cut across the 2 minutes is governed, checked and justified. 7 pairs, 5 questions each, 35 answers written down before the first render. 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 twelve 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. The probability that 8 consecutive shots built from 1 shared header and 1 named reason per cut read as a single continuous sequence is high. The probability that 8 shots assembled from adjectives read as 8 slides is higher than most teams expect, and it costs a full re-render of the sequence to find out. So the method fits on 1 checklist, run top to bottom before any render. 13 rules. 9 slots. 5 questions. 1 clause per cut. Written down once, it stops being an instinct revisited on every shot and becomes a step that anybody on the team can run.