{"course":{"slug":"ai-video","title":"Video with AI","url":"https://addaly.com/learn/ai-video","level":"intermediate","track":"create"},"lesson":{"slug":"what-the-model-ignores","title":"The clauses that quietly do nothing","url":"https://addaly.com/learn/ai-video/what-the-model-ignores","position":29,"of":66,"minutes":9,"answer":"Prompt adherence degrades with the number of independent requirements, and the ones dropped first are counts, spatial relations, negations and anything about timing.","keyIdea":"Prompt adherence degrades with the number of independent requirements, and the ones dropped first are counts, spatial relations, negations and anything about timing, because a few hundred text tokens are steering a quarter of a million video tokens."},"body":{"format":"markdown","text":"## Adherence is a budget, not a guarantee\n\nA prompt with one requirement is usually obeyed. A prompt with six requirements typically gets three or four. This is not the model being lazy; it is what happens when a small conditioning signal has to resolve several constraints simultaneously against a large prior.\n\nWhich ones survive is fairly predictable, and knowing the order lets you put the important requirement where it will not be dropped.\n\n## The reliability order\n\nFrom most to least reliable, on a typical current model:\n\n> Figure: Which clauses survive, and which are quietly dropped. Bars (how often the clause is honoured, per cent): Subject and setting 96; Light direction 88; Shot size and angle 80; One named camera move 72; Subject action as a verb 66; Lens and style 48; Spatial relations 30; Counts 18; Negations 14; Timing in the clip 4. The shape matters, not the numbers; measure your own model by listing your requirements as separate lines and ticking them off against the clip. Everything in the bottom half belongs in the still or in the edit, where it is a fact rather than a request.\n\n1. **Subject identity and setting.** What the thing is and roughly where. Almost always obeyed, and largely supplied by your still anyway.\n2. **Light direction and quality.** Strong, for the reasons in the lighting lesson.\n3. **Shot size and angle.** Strong, though some models reframe less readily from a supplied still.\n4. **Camera move.** Reasonably strong if it is one move, named plainly.\n5. **Subject action as a simple verb.** Turning, walking, sitting, reaching.\n6. **Style and lens character.** Variable by model. Test it.\n7. **Spatial relations between objects.** Weak. \"The cup to the left of the book\" is close to a coin toss.\n8. **Counts.** Very weak. Three of anything is a suggestion.\n9. **Negations.** Very weak, for the reasons in the previous lesson.\n10. **Timing within the clip.** Not available. \"Then she looks up\" has no mechanism behind it.\n\n## The fix is not more words\n\nThe instinct on discovering a dropped requirement is to describe it harder: more adjectives, more emphasis, a longer sentence.\n\nThis usually makes things worse, because adding tokens dilutes every other token's influence. A 120-word prompt is not six times more specific than a 20-word prompt; it is a longer list from which the model picks a similar number of things to honour.\n\nThe things that actually work, in order:\n\n- **Move the requirement into the still.** Counts, spatial relations and object presence all belong in the frame, where they are facts rather than requests. This resolves items 7, 8 and 9 on the list completely.\n- **Split the shot.** If two actions must both happen, that is two shots and a cut. This resolves item 10.\n- **Supply a conditioning signal.** Pose or trajectory for specific movement.\n- **Shorten the prompt.** Cut to the four clauses that matter and let the model handle the rest.\n\n## Diagnosing what was dropped\n\nAfter a disappointing generation, do not rewrite. Do this instead:\n\n1. List your requirements as separate lines.\n2. Watch the clip and tick each one.\n3. Look at the pattern.\n\nIf the dropped ones are counts and spatial relations, that is expected and the fix is the still. If the dropped one is the camera move, try naming it alone in a short prompt to check whether the model knows it at all. If everything was dropped, suspect guidance set too low or a conditioning image that contradicts the text.\n\nThree minutes of this beats twenty minutes of rewriting, and it produces knowledge rather than a new guess.\n\n## The still always wins\n\nWorth stating once more as a standalone rule, because it explains most disappointments.\n\nIf the prompt says \"a red door\" and the still shows a blue door, you get a blue door. The conditioning image is a dense, complete specification of the picture; the text is a few hundred tokens of suggestion. When they disagree, the image wins and the text produces a slight, unhelpful pull toward the thing it named.\n\nSo do not use the prompt to correct the still. Fix the still.\n\n## What this means for how you write\n\nA good video prompt is short, ordered and free of wishes. Something like:\n\n> Wide shot, eye level. A man walks slowly away from camera down a wet alley. Slow dolly in. Hard key from camera right, deep shadows. Everything else static.\n\nFive clauses, all of them in the reliable half of the list. Nothing counted, nothing negated, nothing timed, no spatial relation between objects, no adjective about mood.\n\nEverything that could not be requested reliably has been moved into the still or into the edit, which is where it belonged.\n\n---\n\nToday: take your longest prompt, list its requirements as separate lines, and move every count, negation and spatial relation out of the text and into the frame."},"check":{"question":"A prompt asks for \"three lanterns hanging to the left of the doorway, no people in shot\". The result has five lanterns, mostly on the right, and a figure in the background. What is the correct response?","options":[{"text":"Add emphasis weighting to the lantern count and move the negation to a negative prompt, so both constraints are conditioned more strongly.","correct":false,"tempting":"Applies more pressure to two requirement types that are weak for structural reasons rather than for lack of emphasis."},{"text":"Rewrite the prompt more precisely and at greater length, describing the position of each lantern relative to the doorway and the emptiness of the street.","correct":false,"tempting":"Adds tokens, which dilutes every other clause; length is the opposite of the fix here."},{"text":"Build the lanterns, their positions and the empty street into the first frame, and delete all three requirements from the text.","correct":true},{"text":"Raise the guidance scale, since all three failures are adherence failures and guidance is the control that governs how closely the prompt is followed.","correct":false,"tempting":"Names the adherence control correctly and ignores that counts, spatial relations and negations resist it while motion and colour degrade."}]},"prev":{"slug":"one-variable-at-a-time","title":"An experiment protocol that fits on a card","url":"https://addaly.com/learn/ai-video/one-variable-at-a-time"},"next":{"slug":"consistency-across-shots","title":"A cut breaks on the light before it breaks on the face","url":"https://addaly.com/learn/ai-video/consistency-across-shots"},"authorship":"Founder and series editor: Dr Mustafa Mun. Written by AI models to a written standard. Exam keys checked blind by a second model: 3,665 of 3,666 agreed. Academic review invited.","review":{"status":"Draft, open for academic review"},"cite":"Addaly (2026) 'The clauses that quietly do nothing', in Video with AI. Edited by M. Mun. Addaly. Available at: https://addaly.com/learn/ai-video/what-the-model-ignores (Accessed: 13 October 2026).","licence":{"name":"CC BY-NC-SA 4.0","url":"https://creativecommons.org/licenses/by-nc-sa/4.0/"},"reader":{"schema":1,"title":"The clauses that quietly do nothing","courseTitle":"Video with AI","module":{"index":3,"slug":"directing-the-shot","title":"Directing, not describing"},"answer":null,"keyIdea":"Prompt adherence degrades with the number of independent requirements, and the ones dropped first are counts, spatial relations, negations and anything about timing, because a few hundred text tokens are steering a quarter of a million video tokens.","hero":{"kind":"figure","figure":{"t":"figure","label":"Figure","figure":{"type":"bars","unit":"how often the clause is honoured, per cent","title":"Which clauses survive, and which are quietly dropped","series":[{"label":"Subject and setting","value":96},{"label":"Light direction","value":88},{"label":"Shot size and angle","value":80},{"label":"One named camera move","value":72},{"label":"Subject action as a verb","value":66},{"label":"Lens and style","value":48},{"label":"Spatial relations","value":30},{"label":"Counts","value":18},{"label":"Negations","value":14},{"label":"Timing in the clip","value":4,"highlight":true}],"caption":"The shape matters, not the numbers; measure your own model by listing your requirements as separate lines and ticking them off against the clip. Everything in the bottom half belongs in the still or in the edit, where it is a fact rather than a request."},"alt":"Figure: Which clauses survive, and which are quietly dropped. Bars (how often the clause is honoured, per cent): Subject and setting 96; Light direction 88; Shot size and angle 80; One named camera move 72; Subject action as a verb 66; Lens and style 48; Spatial relations 30; Counts 18; Negations 14; Timing in the clip 4. The shape matters, not the numbers; measure your own model by listing your requirements as separate lines and ticking them off against the clip. Everything in the bottom half belongs in the still or in the edit, where it is a fact rather than a request."}},"blocks":[{"t":"h2","id":"adherence-is-a-budget-not-a-guarantee","runs":[{"text":"Adherence is a budget, not a guarantee"}]},{"t":"p","runs":[{"text":"A prompt with one requirement is usually obeyed. A prompt with six requirements typically gets three or four. This is not the model being lazy; it is what happens when a small conditioning signal has to resolve several constraints simultaneously against a large prior."}]},{"t":"p","runs":[{"text":"Which ones survive is fairly predictable, and knowing the order lets you put the important requirement where it will not be dropped."}]},{"t":"h2","id":"the-reliability-order","runs":[{"text":"The reliability order"}]},{"t":"p","runs":[{"text":"From most to least reliable, on a typical current model:"}]},{"t":"ol","start":1,"items":[[{"text":"Subject identity and setting.","b":true},{"text":" What the thing is and roughly where. Almost always obeyed, and largely supplied by your still anyway."}],[{"text":"Light direction and quality.","b":true},{"text":" Strong, for the reasons in the lighting lesson."}],[{"text":"Shot size and angle.","b":true},{"text":" Strong, though some models reframe less readily from a supplied still."}],[{"text":"Camera move.","b":true},{"text":" Reasonably strong if it is one move, named plainly."}],[{"text":"Subject action as a simple verb.","b":true},{"text":" Turning, walking, sitting, reaching."}],[{"text":"Style and lens character.","b":true},{"text":" Variable by model. Test it."}],[{"text":"Spatial relations between objects.","b":true},{"text":" Weak. “The cup to the left of the book” is close to a coin toss."}],[{"text":"Counts.","b":true},{"text":" Very weak. Three of anything is a suggestion."}],[{"text":"Negations.","b":true},{"text":" Very weak, for the reasons in the previous lesson."}],[{"text":"Timing within the clip.","b":true},{"text":" Not available. “Then she looks up” has no mechanism behind it."}]]},{"t":"h2","id":"the-fix-is-not-more-words","runs":[{"text":"The fix is not more words"}]},{"t":"p","runs":[{"text":"The instinct on discovering a dropped requirement is to describe it harder: more adjectives, more emphasis, a longer sentence."}]},{"t":"p","runs":[{"text":"This usually makes things worse, because adding tokens dilutes every other token’s influence. A 120-word prompt is not six times more specific than a 20-word prompt; it is a longer list from which the model picks a similar number of things to honour."}]},{"t":"p","runs":[{"text":"The things that actually work, in order:"}]},{"t":"ul","items":[[{"text":"Move the requirement into the still.","b":true},{"text":" Counts, spatial relations and object presence all belong in the frame, where they are facts rather than requests. This resolves items 7, 8 and 9 on the list completely."}],[{"text":"Split the shot.","b":true},{"text":" If two actions must both happen, that is two shots and a cut. This resolves item 10."}],[{"text":"Supply a conditioning signal.","b":true},{"text":" Pose or trajectory for specific movement."}],[{"text":"Shorten the prompt.","b":true},{"text":" Cut to the four clauses that matter and let the model handle the rest."}]]},{"t":"h2","id":"diagnosing-what-was-dropped","runs":[{"text":"Diagnosing what was dropped"}]},{"t":"p","runs":[{"text":"After a disappointing generation, do not rewrite. Do this instead:"}]},{"t":"ol","start":1,"items":[[{"text":"List your requirements as separate lines."}],[{"text":"Watch the clip and tick each one."}],[{"text":"Look at the pattern."}]]},{"t":"p","runs":[{"text":"If the dropped ones are counts and spatial relations, that is expected and the fix is the still. If the dropped one is the camera move, try naming it alone in a short prompt to check whether the model knows it at all. If everything was dropped, suspect guidance set too low or a conditioning image that contradicts the text."}]},{"t":"p","runs":[{"text":"Three minutes of this beats twenty minutes of rewriting, and it produces knowledge rather than a new guess."}]},{"t":"h2","id":"the-still-always-wins","runs":[{"text":"The still always wins"}]},{"t":"p","runs":[{"text":"Worth stating once more as a standalone rule, because it explains most disappointments."}]},{"t":"p","runs":[{"text":"If the prompt says “a red door” and the still shows a blue door, you get a blue door. The conditioning image is a dense, complete specification of the picture; the text is a few hundred tokens of suggestion. When they disagree, the image wins and the text produces a slight, unhelpful pull toward the thing it named."}]},{"t":"p","runs":[{"text":"So do not use the prompt to correct the still. Fix the still."}]},{"t":"h2","id":"what-this-means-for-how-you-write","runs":[{"text":"What this means for how you write"}]},{"t":"p","runs":[{"text":"A good video prompt is short, ordered and free of wishes. Something like:"}]},{"t":"quote","runs":[{"text":"Wide shot, eye level. A man walks slowly away from camera down a wet alley. Slow dolly in. Hard key from camera right, deep shadows. Everything else static."}]},{"t":"p","runs":[{"text":"Five clauses, all of them in the reliable half of the list. Nothing counted, nothing negated, nothing timed, no spatial relation between objects, no adjective about mood."}]},{"t":"p","runs":[{"text":"Everything that could not be requested reliably has been moved into the still or into the edit, which is where it belonged."}]},{"t":"rule"},{"t":"p","runs":[{"text":"Today: take your longest prompt, list its requirements as separate lines, and move every count, negation and spatial relation out of the text and into the frame."}]}],"check":{"question":[{"text":"A prompt asks for “three lanterns hanging to the left of the doorway, no people in shot”. The result has five lanterns, mostly on the right, and a figure in the background. What is the correct response?"}],"options":[{"text":[{"text":"Add emphasis weighting to the lantern count and move the negation to a negative prompt, so both constraints are conditioned more strongly."}],"correct":false,"tempting":[{"text":"Applies more pressure to two requirement types that are weak for structural reasons rather than for lack of emphasis."}]},{"text":[{"text":"Rewrite the prompt more precisely and at greater length, describing the position of each lantern relative to the doorway and the emptiness of the street."}],"correct":false,"tempting":[{"text":"Adds tokens, which dilutes every other clause; length is the opposite of the fix here."}]},{"text":[{"text":"Build the lanterns, their positions and the empty street into the first frame, and delete all three requirements from the text."}],"correct":true,"tempting":null},{"text":[{"text":"Raise the guidance scale, since all three failures are adherence failures and guidance is the control that governs how closely the prompt is followed."}],"correct":false,"tempting":[{"text":"Names the adherence control correctly and ignores that counts, spatial relations and negations resist it while motion and colour degrade."}]}],"order":[3,0,2,1],"why":null},"prev":{"slug":"one-variable-at-a-time","title":"An experiment protocol that fits on a card","minutes":8},"next":{"slug":"consistency-across-shots","title":"A cut breaks on the light before it breaks on the face","minutes":10}}}