{"course":{"slug":"ai-images","title":"Image Generation, in Practice","url":"https://addaly.com/learn/ai-images","level":"beginner","track":"create"},"lesson":{"slug":"the-prompt-in-slots","title":"Build the prompt in slots, not as a sentence you keep rewriting","url":"https://addaly.com/learn/ai-images/the-prompt-in-slots","position":22,"of":69,"minutes":8,"answer":"Writing a prompt as seven named slots makes it debuggable, because when the image is wrong you can change one slot and know which part of the description caused the change.","keyIdea":"Writing a prompt as seven named slots makes it debuggable, because when the image is wrong you can change one slot and know which part of the description caused the change."},"body":{"format":"markdown","text":"## Why prompts get worse as you edit them\n\nWatch somebody work on a prompt for twenty minutes and you see the same pattern. They add a word, generate, do not like it, add three more, generate, delete something, add \"highly detailed\", generate. After twenty minutes the prompt is ninety words long, nobody can say what any of it is doing, and the images are not better than attempt four.\n\nThe problem is that the prompt has no structure, so there is nothing to change *one of*. The fix is to give it slots.\n\n## Seven slots\n\nWrite the prompt as an ordered list of named parts. Keep the names in your notes, even if the model only ever sees the joined text.\n\n1. **Medium**: photograph, oil painting, technical illustration, 3D render\n2. **Subject**: who or what, with the two or three attributes that matter\n3. **Action or state**: what they are doing, or explicitly at rest\n4. **Setting**: where, and how much of it is visible\n5. **Light**: direction, quality, time of day, colour\n6. **Camera**: framing, focal length, height, depth of field\n7. **Finish**: grain, film stock, process, colour treatment\n\nA worked example:\n\n```text\nmedium:  documentary photograph\nsubject: an older woman in a green cotton sari, silver bangles\naction:  sorting dried chillies on a low table, hands in motion\nsetting: an open courtyard, whitewashed wall behind, few props\nlight:   hard late-afternoon sun from the left, deep shadows\ncamera:  waist-level, 35mm, full body with headroom, slight wide angle\nfinish:  colour negative film, fine grain, slightly cool shadows\n```\n\nJoined with commas, that is a prompt. Kept as a list, it is a specification you can debug.\n\n> Figure: Seven named slots, so there is something to change one of. Layers, top to bottom: Medium (Photograph, oil painting, technical illustration, 3D render); Subject (Who or what, with the two or three attributes that matter); Action or state (What they are doing, or explicitly at rest); Setting (Where, and how much of it is visible); Light (Direction, quality, time of day, colour); Camera (Framing, focal length, height, depth of field); Finish (Grain, film stock, process, colour treatment). Light is the slot missing most often, and the one whose absence makes an image look generated. Slots five to seven lift straight out into the next job as a house look.\n\n## What the slots buy you\n\n**Coverage.** Most weak prompts are missing a slot entirely, and light is missing most often. Left unspecified, the model gives you its average, which on photographic checkpoints means the soft, even, everywhere-at-once light that makes generated pictures recognisable as generated.\n\n**Debugging.** The image is too tight. That is the camera slot, not the subject. The colours are wrong. That is finish or light. Without slots, people respond to a framing problem by changing the subject description, and then cannot explain why it got worse.\n\n**Reuse.** Slots 5, 6 and 7 are the visual system from your brief. Fix them across twelve images and you have the consistency the set block will ask for, at no extra cost.\n\n## Order, and how much it matters\n\nIt matters, but less than it used to, and it matters differently per architecture.\n\nCLIP-based models, SD 1.5, SDXL, read the prompt in 75-token chunks and weight earlier tokens somewhat more heavily. Put the subject early there.\n\nT5-based models, the FLUX family and several newer systems, handle longer natural-language descriptions and are far less position-sensitive. With those, full sentences outperform comma-separated tags, and burying an important detail at the end is much less punishing.\n\nThe practical rule that survives both: **put the thing the image is about in the first fifteen words, and do not rely on anything after word sixty.** Which brings us to the real ceiling.\n\n## The limit nobody tells you\n\nAdding more words stops working, and the mechanism is worth knowing. CLIP text encoders have a 77-token window. Interfaces get around it by splitting longer prompts into chunks, encoding each separately, and concatenating the results. There is no attention across the chunk boundary, so a relationship you describe in words 80 to 90 cannot bind to a subject named in word 5.\n\nThat is why enormous prompts produce images containing all the right *things* in none of the right relationships. The model is not ignoring you; the architecture has no path for that particular instruction to reach that particular subject.\n\nSo when a prompt passes about sixty words and is still not working, stop adding. Either move the requirement to a different mechanism (a reference image, a control, an inpainting pass) or accept that this part gets fixed in editing. The next block is entirely about that first option, and it is where prompts stop being the main tool.\n\n## Keep the slots as a file\n\nThe practical form of this is a small text file per project rather than a prompt box you type into.\n\n```text\n# courtyard hero — slots\nMEDIUM   documentary photograph\nSUBJECT  older woman, green cotton sari, silver bangles\nACTION   sorting dried chillies on a low table, hands in motion\nSETTING  open courtyard, whitewashed wall, few props\nLIGHT    hard late-afternoon sun from the left, deep shadows\nCAMERA   waist level, 35mm, full body with headroom\nFINISH   colour negative film, fine grain, cool shadows\n```\n\nThree things follow from having it on disk. You can diff two versions and see exactly what changed between Tuesday and Thursday. You can hand it to somebody else and they can work from it. And slots 5 to 7 lift straight out into the next job as a house look, which is where a personal style starts to come from.\n\nIt also stops the slow creep. When the prompt lives in a box you keep editing, it only ever grows. When it lives in named fields, adding a word means deciding which field it belongs to, and about a third of the time you realise it belongs to none of them and should not be there."},"check":{"question":"A ninety-word prompt describes a woman in a blue coat and a man in a red scarf, and the model keeps swapping the colours. Adding emphasis does not fix it. What is the mechanism?","options":[{"text":"Colour words are weakly represented in the text encoder, so they bind to the scene as a whole rather than to individual subjects mentioned in the prompt.","correct":false,"tempting":"Colour binds tightly enough in short prompts, which is the clue that length rather than the colour vocabulary is what broke it."},{"text":"Attribute binding weakens across the 77-token chunk boundary, so a long prompt loses the link between a subject and its attributes.","correct":true},{"text":"The guidance scale is too low for the model to enforce two separate attribute assignments within one image at the same time.","correct":false,"tempting":"Guidance strengthens adherence to the prompt as a whole; raising it on a prompt with broken binding produces a more emphatic version of the same swap."},{"text":"The two subjects are too similar in the training distribution, so the model treats them as one concept with two interchangeable sets of attributes.","correct":false,"tempting":"A real effect for genuinely similar subjects, but it would not depend on prompt length, and the same pair described briefly usually binds correctly."}]},"prev":{"slug":"rights-and-disclosure","title":"What you can sell, and what you must disclose","url":"https://addaly.com/learn/ai-images/rights-and-disclosure"},"next":{"slug":"lens-and-camera-words","title":"The camera vocabulary that works, and the words that do nothing","url":"https://addaly.com/learn/ai-images/lens-and-camera-words"},"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) 'Build the prompt in slots, not as a sentence you keep rewriting', in Image Generation, in Practice. Edited by M. Mun. Addaly. Available at: https://addaly.com/learn/ai-images/the-prompt-in-slots (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":"Build the prompt in slots, not as a sentence you keep rewriting","courseTitle":"Image Generation, in Practice","module":{"index":3,"slug":"first-frames","title":"Getting to a frame worth keeping"},"answer":null,"keyIdea":"Writing a prompt as seven named slots makes it debuggable, because when the image is wrong you can change one slot and know which part of the description caused the change.","hero":{"kind":"figure","figure":{"t":"figure","label":"Figure 3.1","figure":{"type":"stack","title":"Seven named slots, so there is something to change one of","layers":[{"label":"Medium","detail":"Photograph, oil painting, technical illustration, 3D render"},{"label":"Subject","detail":"Who or what, with the two or three attributes that matter"},{"label":"Action or state","detail":"What they are doing, or explicitly at rest"},{"label":"Setting","detail":"Where, and how much of it is visible"},{"label":"Light","detail":"Direction, quality, time of day, colour"},{"label":"Camera","detail":"Framing, focal length, height, depth of field"},{"label":"Finish","detail":"Grain, film stock, process, colour treatment"}],"caption":"Light is the slot missing most often, and the one whose absence makes an image look generated. Slots five to seven lift straight out into the next job as a house look."},"alt":"Figure: Seven named slots, so there is something to change one of. Layers, top to bottom: Medium (Photograph, oil painting, technical illustration, 3D render); Subject (Who or what, with the two or three attributes that matter); Action or state (What they are doing, or explicitly at rest); Setting (Where, and how much of it is visible); Light (Direction, quality, time of day, colour); Camera (Framing, focal length, height, depth of field); Finish (Grain, film stock, process, colour treatment). Light is the slot missing most often, and the one whose absence makes an image look generated. Slots five to seven lift straight out into the next job as a house look."}},"blocks":[{"t":"h2","id":"why-prompts-get-worse-as-you-edit-them","runs":[{"text":"Why prompts get worse as you edit them"}]},{"t":"p","runs":[{"text":"Watch somebody work on a prompt for twenty minutes and you see the same pattern. They add a word, generate, do not like it, add three more, generate, delete something, add “highly detailed”, generate. After twenty minutes the prompt is ninety words long, nobody can say what any of it is doing, and the images are not better than attempt four."}]},{"t":"p","runs":[{"text":"The problem is that the prompt has no structure, so there is nothing to change "},{"text":"one of","i":true},{"text":". The fix is to give it slots."}]},{"t":"h2","id":"seven-slots","runs":[{"text":"Seven slots"}]},{"t":"p","runs":[{"text":"Write the prompt as an ordered list of named parts. Keep the names in your notes, even if the model only ever sees the joined text."}]},{"t":"ol","start":1,"items":[[{"text":"Medium","b":true},{"text":": photograph, oil painting, technical illustration, 3D render"}],[{"text":"Subject","b":true},{"text":": who or what, with the two or three attributes that matter"}],[{"text":"Action or state","b":true},{"text":": what they are doing, or explicitly at rest"}],[{"text":"Setting","b":true},{"text":": where, and how much of it is visible"}],[{"text":"Light","b":true},{"text":": direction, quality, time of day, colour"}],[{"text":"Camera","b":true},{"text":": framing, focal length, height, depth of field"}],[{"text":"Finish","b":true},{"text":": grain, film stock, process, colour treatment"}]]},{"t":"p","runs":[{"text":"A worked example:"}]},{"t":"code","lang":"text","label":"Text","text":"medium:  documentary photograph\nsubject: an older woman in a green cotton sari, silver bangles\naction:  sorting dried chillies on a low table, hands in motion\nsetting: an open courtyard, whitewashed wall behind, few props\nlight:   hard late-afternoon sun from the left, deep shadows\ncamera:  waist-level, 35mm, full body with headroom, slight wide angle\nfinish:  colour negative film, fine grain, slightly cool shadows"},{"t":"p","runs":[{"text":"Joined with commas, that is a prompt. Kept as a list, it is a specification you can debug."}]},{"t":"h2","id":"what-the-slots-buy-you","runs":[{"text":"What the slots buy you"}]},{"t":"p","runs":[{"text":"Coverage.","b":true},{"text":" Most weak prompts are missing a slot entirely, and light is missing most often. Left unspecified, the model gives you its average, which on photographic checkpoints means the soft, even, everywhere-at-once light that makes generated pictures recognisable as generated."}]},{"t":"p","runs":[{"text":"Debugging.","b":true},{"text":" The image is too tight. That is the camera slot, not the subject. The colours are wrong. That is finish or light. Without slots, people respond to a framing problem by changing the subject description, and then cannot explain why it got worse."}]},{"t":"p","runs":[{"text":"Reuse.","b":true},{"text":" Slots 5, 6 and 7 are the visual system from your brief. Fix them across twelve images and you have the consistency the set block will ask for, at no extra cost."}]},{"t":"h2","id":"order-and-how-much-it-matters","runs":[{"text":"Order, and how much it matters"}]},{"t":"p","runs":[{"text":"It matters, but less than it used to, and it matters differently per architecture."}]},{"t":"p","runs":[{"text":"CLIP-based models, SD 1.5, SDXL, read the prompt in 75-token chunks and weight earlier tokens somewhat more heavily. Put the subject early there."}]},{"t":"p","runs":[{"text":"T5-based models, the FLUX family and several newer systems, handle longer natural-language descriptions and are far less position-sensitive. With those, full sentences outperform comma-separated tags, and burying an important detail at the end is much less punishing."}]},{"t":"p","runs":[{"text":"The practical rule that survives both: "},{"text":"put the thing the image is about in the first fifteen words, and do not rely on anything after word sixty.","b":true},{"text":" Which brings us to the real ceiling."}]},{"t":"h2","id":"the-limit-nobody-tells-you","runs":[{"text":"The limit nobody tells you"}],"role":"limit"},{"t":"p","runs":[{"text":"Adding more words stops working, and the mechanism is worth knowing. CLIP text encoders have a 77-token window. Interfaces get around it by splitting longer prompts into chunks, encoding each separately, and concatenating the results. There is no attention across the chunk boundary, so a relationship you describe in words 80 to 90 cannot bind to a subject named in word 5."}]},{"t":"p","runs":[{"text":"That is why enormous prompts produce images containing all the right "},{"text":"things","i":true},{"text":" in none of the right relationships. The model is not ignoring you; the architecture has no path for that particular instruction to reach that particular subject."}]},{"t":"p","runs":[{"text":"So when a prompt passes about sixty words and is still not working, stop adding. Either move the requirement to a different mechanism (a reference image, a control, an inpainting pass) or accept that this part gets fixed in editing. The next block is entirely about that first option, and it is where prompts stop being the main tool."}]},{"t":"h2","id":"keep-the-slots-as-a-file","runs":[{"text":"Keep the slots as a file"}]},{"t":"p","runs":[{"text":"The practical form of this is a small text file per project rather than a prompt box you type into."}]},{"t":"code","lang":"text","label":"Text","text":"# courtyard hero — slots\nMEDIUM   documentary photograph\nSUBJECT  older woman, green cotton sari, silver bangles\nACTION   sorting dried chillies on a low table, hands in motion\nSETTING  open courtyard, whitewashed wall, few props\nLIGHT    hard late-afternoon sun from the left, deep shadows\nCAMERA   waist level, 35mm, full body with headroom\nFINISH   colour negative film, fine grain, cool shadows"},{"t":"p","runs":[{"text":"Three things follow from having it on disk. You can diff two versions and see exactly what changed between Tuesday and Thursday. You can hand it to somebody else and they can work from it. And slots 5 to 7 lift straight out into the next job as a house look, which is where a personal style starts to come from."}]},{"t":"p","runs":[{"text":"It also stops the slow creep. When the prompt lives in a box you keep editing, it only ever grows. When it lives in named fields, adding a word means deciding which field it belongs to, and about a third of the time you realise it belongs to none of them and should not be there."}]}],"check":{"question":[{"text":"A ninety-word prompt describes a woman in a blue coat and a man in a red scarf, and the model keeps swapping the colours. Adding emphasis does not fix it. What is the mechanism?"}],"options":[{"text":[{"text":"Colour words are weakly represented in the text encoder, so they bind to the scene as a whole rather than to individual subjects mentioned in the prompt."}],"correct":false,"tempting":[{"text":"Colour binds tightly enough in short prompts, which is the clue that length rather than the colour vocabulary is what broke it."}]},{"text":[{"text":"Attribute binding weakens across the 77-token chunk boundary, so a long prompt loses the link between a subject and its attributes."}],"correct":true,"tempting":null},{"text":[{"text":"The guidance scale is too low for the model to enforce two separate attribute assignments within one image at the same time."}],"correct":false,"tempting":[{"text":"Guidance strengthens adherence to the prompt as a whole; raising it on a prompt with broken binding produces a more emphatic version of the same swap."}]},{"text":[{"text":"The two subjects are too similar in the training distribution, so the model treats them as one concept with two interchangeable sets of attributes."}],"correct":false,"tempting":[{"text":"A real effect for genuinely similar subjects, but it would not depend on prompt length, and the same pair described briefly usually binds correctly."}]}],"order":[3,2,1,0],"why":null},"prev":{"slug":"rights-and-disclosure","title":"What you can sell, and what you must disclose","minutes":10},"next":{"slug":"lens-and-camera-words","title":"The camera vocabulary that works, and the words that do nothing","minutes":8}}}