The 180-degree rule and the eyeline match don’t disappear just because your camera is a prompt, not a machine on a dolly.I generated two shots for a dialogue scene in Lost Garden last month, cut them together, and watched two characters who were supposed to be facing each other end up staring in the same direction, like strangers waiting for the same bus. Nothing was wrong with either shot on its own. Both were clean, well lit, on model. Cut together, the scene told the audience the two people weren’t looking at each other at all.Keeping screen direction consistent in AI-generated video means deciding, before you generate anything, which side of the action a shot sits on, then writing that decision down somewhere more permanent than your memory. AI video tools have no camera object that persists between generations. Nothing enforces the line for you. That job now belongs entirely to you, and it’s easy to lose track of when every shot is its own isolated request to a model that has never seen the shot before it.This is the workflow I run to keep it from happening again.What is screen direction, and why does AI video break it by default?Screen direction is the consistent left-right relationship between characters and their eyelines across a cut. If a character looks screen-left at another character in one shot, they need to keep looking screen-left at them in the next shot of that exchange, or the geography of the scene falls apart in the viewer’s head.On a real set, this is enforced by the camera itself. The operator’s rig physically sits on one side of an imaginary line drawn between the two subjects, called the axis of action, and every new setup gets checked against where the last one was.An AI video generator has no rig, no set, and no memory of the last shot. Each generation is a fresh request built from a text prompt and maybe a reference image. The model doesn’t know a person was standing on the left in the previous clip unless you tell it, explicitly, every single time. Skip that instruction and the model defaults to whatever composition looks best for that one prompt in isolation, which is exactly how you end up with two people who are supposedly talking to each other both facing the same direction.Why does the 180-degree rule still apply when there’s no camera to move?The 180-degree rule isn’t a hardware constraint. It’s a convention about audience orientation that predates any specific camera at all. Continuity editing, including the 180-degree rule and the eyeline match, took shape in Hollywood during the 1910s and 1920s, with D.W. Griffith and Edwin S. Porter credited among the filmmakers who established the core techniques, and Soviet filmmaker Lev Kuleshov’s editing experiments in the same decade reinforcing how much spatial meaning a cut can carry on its own. The rule survived nearly a century of new formats, new cameras, and new screens because it was never really about equipment. It’s about the audience’s ability to build a stable mental map of where people are standing and which way they’re looking.That’s why the rule doesn’t get suspended just because the “camera” generating your shot is a diffusion model instead of a Panavision body. The audience’s brain doing the orienting hasn’t changed. Only the thing responsible for respecting the line has changed, and right now that thing is you, not the software.The line never lived in the camera. It lived in the director’s head, and someone had to remember to check it before every setup. AI video removed the camera. It didn’t remove the checking.How do you track the line and eyelines across shots from different tools?This gets harder the moment your shots come from more than one tool or more than one session, which for most of us is constantly. A shot generated in Runway on Monday and a shot generated in Kling on Wednesday share nothing: no project file, no camera rig, no institutional memory of what happened in the last clip. Each tool only knows what’s in the prompt and the reference image you feed it that day.So the log has to live outside both tools. What I keep, scene by scene, before I generate a single frame:The axis of action for the scene, described in plain language (“the line runs between the two characters at the kitchen table, camera stays on the window side”).Each character’s eyeline direction for every shot in the sequence, screen-left or screen-right, not “looking at the other person” (too vague to catch a drift).Which side of the line the previous shot was generated from, so the next prompt can say “same side” or flag an intentional cross.A thumbnail or frame grab of the last generated shot, attached to the log entry, because reading a description is slower than glancing at the actual composition.A flag for any deliberate line crossing, with the reason, so it reads as a choice later and not a mistake you’re trying to explain away.A one-line note in this log takes fifteen seconds to write and saves you from a full regeneration later, which, at current model pricing and queue times, is never a fifteen-second cost.What’s the actual step-by-step workflow?Lock the line before you write a single prompt. Decide where the axis of action sits for the scene, based on the blocking you want, not on what’s convenient to generate.Write the eyeline for every character into the shot description, not just “talking to each other.” Screen-left, screen-right, up, down: pick the actual direction and put it in the prompt.Generate the first shot, then check it against the log, not against your memory of what you meant to ask for.Carry the last frame forward as a reference image when the tool supports it, so the next generation has something concrete to match instead of a text description alone.If you have to cross the line, do it on purpose, either with a neutral shot taken roughly on the axis itself, or with a beat that clearly signals the perspective shift, and log it as a decision.Skipping step one is the mistake I see most, including in my own early Lost Garden shots: generating first and checking continuity in the edit, when a five-minute decision made before generation would have caught it for free.Where does this actually go wrong in practice?A few patterns show up often enough to name directly:Reusing a character reference image without checking its facing direction. The reference was shot facing screen-right for a different scene, and it quietly imports that direction into a scene where the character needs to face screen-left.Switching tools mid-scene without restating the line. Each tool starts from zero. A workflow that works inside one tool’s session memory falls apart the moment you switch to a second tool for a single shot.Trusting a wide establishing shot to set the geography, then never checking coverage against it. The wide shot might be perfect and every closer shot afterward can still drift, because nothing connects them except your attention.This is the part of AI filmmaking that has nothing to do with prompt engineering and everything to do with information you have to maintain by hand, on top of the generation itself. It’s also where I lean on ScreenWeaver for my own work: the storyboard step is where I write the axis of action and each character’s eyeline into the shot plan before a single frame gets generated, so the decision exists as a document I can check, not a thing I’m trying to hold in my head across forty shots.FAQCan you fix a crossed line after the shots are already generated?Sometimes, with an establishing or neutral shot inserted between the two conflicting angles to reset the audience’s orientation, but it’s a patch, not a fix. It’s cheaper to catch it in the log before generating the second shot.Do any AI video tools track screen direction automatically?Not reliably. Reference images and character-consistency features help keep a face or outfit stable, but they don’t reason about which side of an axis of action a shot belongs on. That judgment call is still yours.Does this only matter for dialogue scenes?No. Any sequence where the audience needs to track spatial relationships, a chase, a fight, two people walking toward or away from each other, depends on the same discipline. Dialogue just makes the failure easiest to spot.Screen direction was never free, even on a real set with a real camera. It required someone to check it every time. AI video didn’t remove that job. It just removed the machine that used to remind you the job existed.If you’re generating a multi-shot scene this week, write down where your line sits before you write your first prompt. It’s the cheapest insurance in the whole workflow.Sources: Adobe, “What is the 180-degree rule?”; Wikipedia, “Continuity editing.”