Image Generation prompts

Help me write an image generation prompt.

WHAT I WANT TO SEE: [DESCRIBE IT, HOWEVER ROUGHLY]
WHAT IT IS FOR: [website hero / product shot / illustration / social post / concept art / personal]
MODEL: [Midjourney / DALL-E / Stable Diffusion / Imagen / other]
ASPECT RATIO: [DIMENSIONS]
WHAT I HAVE TRIED ALREADY: [PREVIOUS PROMPTS AND WHAT WENT WRONG]

Build the prompt in layers, and show each layer separately so I can adjust one at a time:

1. SUBJECT - what is in the frame, specifically. The single most common failure is a vague subject. 'A woman' produces anything; 'a woman in her sixties in a hospital porter's uniform, mid-stride' produces something.
2. ACTION AND POSE - what they are doing, where they are looking.
3. SETTING - where, including what is behind and around.
4. COMPOSITION - shot type (close-up, medium, wide), camera angle, where the subject sits in the frame, depth of field, and what is in focus.
5. LIGHTING - direction, quality (hard or soft), time of day, colour temperature, and the source. Lighting does more for the feel of an image than any style keyword.
6. STYLE MEDIUM - photograph, oil painting, line drawing, 3D render, and the specific qualities of that medium.
7. COLOUR - the palette, in words.
8. MOOD - the feeling, stated once, not repeatedly.

Then produce:
- THE ASSEMBLED PROMPT for my model, in its conventions
- WHAT TO PUT IN THE NEGATIVE PROMPT if my model supports one
- THREE VARIATIONS changing one layer each, so I can see what each layer controls
- WHAT WILL PROBABLY GO WRONG - the specific elements models handle badly: hands, text, precise counts of objects, exact spatial relationships, and consistent details across a set
- WHAT TO CHANGE FIRST if the result is not right

Rules:
- Do not stack style keywords hoping something lands. Each word should do a job.
- No named living artists' styles.
- Describe what you want present, not what you want absent, except in a negative prompt field.

My image prompt is not producing what I want. Diagnose it.

MY PROMPT:
"""
[PASTE]
"""

WHAT I WANTED: [DESCRIPTION]
WHAT I AM GETTING: [DESCRIPTION OF THE ACTUAL OUTPUT]
MODEL AND SETTINGS: [MODEL, VERSION, ANY PARAMETERS]
HOW MANY ATTEMPTS: [COUNT]

Work through the usual causes:

1. THE VAGUE SUBJECT - is the main subject described specifically enough? Quote the vague parts.

2. CONFLICTING INSTRUCTIONS - terms pulling in different directions. 'Minimalist' and 'ornate', 'candid' and 'perfectly composed', a style keyword that contradicts the medium. Models resolve conflicts unpredictably.

3. TOO MANY ELEMENTS - the more you ask for in one image, the less control you have over each. Count the distinct requirements in my prompt; over about five or six, something will be dropped. Say which are most likely to be ignored.

4. WORD ORDER AND WEIGHT - most models weight earlier terms more heavily. Is the thing I care about most at the front? For models supporting weights, where to apply them.

5. THE ABSTRACT WORD PROBLEM - words like 'professional', 'modern', 'high quality' and 'beautiful' carry no visual information and consume attention. Identify them in my prompt and replace each with a concrete visual description.

6. THE NEGATIVE SPACE PROBLEM - describing what you do not want in the main prompt often summons it. Move those to the negative prompt or rephrase positively.

7. MODEL LIMITS - the things this model simply does poorly: readable text, hands and fingers, exact object counts, specific spatial arrangements, consistent characters across images, and anything requiring the model to count or do arithmetic. Say whether my request runs into one of these, because no prompt change will fix it.

8. SETTINGS - whether a parameter is working against me: stylisation, guidance scale, aspect ratio affecting composition, or a seed producing repeated results.

Then:
- THE DIAGNOSIS - the single main problem
- THE REWRITTEN PROMPT
- THE THREE-STEP TEST - strip the prompt to its core, confirm that works, then add elements back one at a time. This isolates the offending term.
- WHAT TO DO IF THE MODEL SIMPLY CANNOT DO THIS - compose it in parts and combine them, or edit afterwards.

Help me generate a consistent set of images.

WHAT NEEDS TO BE CONSISTENT: [a character / a style / a product / a setting]
THE SUBJECT: [DESCRIBE IT IN DETAIL]
HOW MANY IMAGES: [COUNT]
WHAT VARIES BETWEEN THEM: [poses, scenes, angles, products]
MODEL: [WHICH ONE, AND WHAT CONSISTENCY FEATURES IT OFFERS - reference images, character references, seeds, LoRAs]
WHAT IT IS FOR: [PURPOSE]

Produce:

1. THE HONEST CAPABILITY CHECK first. Consistency across generated images is genuinely hard and most models do it imperfectly. State what my model can and cannot hold steady, and how much variation to expect. If the use needs exact consistency - a brand mascot, a product's actual appearance - say plainly that generation may be the wrong tool and what to do instead.

2. THE CONSISTENCY ANCHOR - the fixed description that appears identically in every prompt. Write it once, precisely, covering only the features that must not change. Keep it short; a long anchor competes with the varying elements.

3. THE VARIABLE SECTION - the template for what changes per image.

4. THE FULL PROMPT SET - all [COUNT] prompts, each combining the anchor with its variation. Identical wording and order for the anchor in every one; even small rewordings shift the result.

5. THE TECHNICAL LEVERS for my model - reference or character-reference images, seed control, style references, or fine-tuning. Say which are available, what each does, and which to use first.

6. WHAT WILL DRIFT ANYWAY - the features that vary between generations no matter what: fine facial details, exact clothing details, small props, and background specifics. Plan around them rather than fighting them.

7. THE WORKFLOW - generate the anchor image first, pick the best, then use it as a reference for the rest. This is far more reliable than generating the set from text alone.

8. THE SELECTION STRATEGY - generate more than you need and select for consistency. Expect a meaningful reject rate and factor it into the effort estimate.

9. THE POST-PROCESSING - what to fix afterwards in an editor rather than trying to prompt perfectly: colour matching across the set, cropping to consistent composition, and small detail corrections.

Help me achieve a specific visual style.

THE LOOK I WANT: [DESCRIBE IT - reference a period, a medium, a technique, a feeling]
WHAT IT IS FOR: [PURPOSE AND WHERE IT APPEARS]
MODEL: [WHICH]
WHAT I AM GETTING INSTEAD: [THE DEFAULT LOOK YOU WANT TO ESCAPE]

Produce:

1. THE STYLE BROKEN DOWN - any visual style is a set of specific choices. Decompose what I described into:
   - MEDIUM AND TECHNIQUE: what it was physically made with, and the marks that medium leaves
   - LIGHTING: direction, hardness, colour, contrast ratio
   - COLOUR: palette, saturation, where the darks and lights sit
   - COMPOSITION: framing conventions, where subjects are placed, negative space
   - TEXTURE AND SURFACE: grain, brushwork, imperfections
   - SUBJECT TREATMENT: how people or objects are typically depicted in this style

2. THE VOCABULARY - the specific terms that describe each of these, rather than the style's name. Naming a style produces the model's cliché of it; describing its components produces the style.

3. THE PROMPT - built from the decomposition.

4. THE ANTI-DEFAULT TERMS - the generated look I said I want to escape comes from specific defaults: excessive smoothness, symmetrical faces, over-saturated colour, dramatic rim lighting, and shallow depth of field everywhere. Name the terms that counteract each, and what to put in a negative prompt.

5. THE IMPERFECTIONS - most convincing styles include flaws the model will smooth away: grain, uneven lighting, awkward framing, dust, chromatic aberration, blur from movement. Name the ones appropriate here and how to ask for them.

6. THE TECHNICAL SPECIFICS - if the style has real-world equivalents, the terminology that carries them: film stocks, lens characteristics, printing processes, paint types, paper stock. These terms encode a great deal of visual information in few words.

7. THE TEST SET - five prompts varying one element each, so I can find which terms are doing the work.

Rules:
- Do not use living artists' names.
- Describe the style's properties, not a shortcut to it.
- One style. Mixing three produces a muddle.

Help me think through using generated images commercially.

WHAT I WANT TO GENERATE: [DESCRIPTION]
COMMERCIAL USE: [advertising / product packaging / website / print / merchandise / editorial]
MODEL AND PLAN: [WHICH SERVICE AND WHAT TIER]
JURISDICTION: [WHERE YOU OPERATE]

Produce:

1. THE PROMPT ITSELF - what to avoid asking for, and why:
   - Named living artists' styles
   - Recognisable characters, mascots, logos, or brand designs
   - Identifiable real people
   - Recreations of specific existing artworks, album covers, posters or film stills
   - Trade dress: a product design closely associated with a brand
   Say which of these my described image risks, and how to rephrase it to get what I actually need without them.

2. WHAT THE OUTPUT MIGHT ACCIDENTALLY CONTAIN - generated images sometimes include distorted logos, text resembling brand marks, or faces resembling real people. For commercial use, someone must actually look at each image for these. Give the checklist.

3. THE TERMS QUESTION - services differ on commercial rights, and by plan tier. State that I need to read my specific service's current terms, and list the specific questions to answer: does my tier permit commercial use, who owns the output, are there attribution requirements, and are there restrictions by use type.

4. THE COPYRIGHT UNCERTAINTY - be honest that whether purely generated images attract copyright protection varies by jurisdiction and is unsettled in several. The practical implication: you may not be able to stop someone else using your generated image. This matters for a logo or a brand asset and matters less for a blog header.

5. THE USE-TYPE RISK RANKING - for my stated use, where it sits. A blog illustration is low risk; a logo, packaging, or a national advertising campaign is high risk and warrants real legal review.

6. WHERE GENERATION IS THE WRONG CHOICE - for anything that must be exclusively yours, must be trademarked, or must depict a real product accurately. Say so plainly if my use falls here, and name the alternative.

7. DISCLOSURE - whether the context expects you to say an image is AI-generated: editorial, journalism, some advertising standards, and some platforms.

8. THE PRACTICAL CHECKLIST before publishing any generated image commercially.

I am not a lawyer and this is not legal advice. For high-risk uses, say clearly that a lawyer should review it.

Help me plan a set of images.

THE PROJECT: [WHAT IT IS - website, article series, deck, product listing, campaign]
HOW MANY IMAGES: [COUNT AND WHERE EACH GOES]
WHAT EACH NEEDS TO COMMUNICATE: [LIST BY SLOT]
OVERALL FEELING: [THE TONE]
CONSTRAINTS: [dimensions, file sizes, whether text overlays go on top]
MODEL: [WHICH]

Produce:

1. THE VISUAL SYSTEM - the decisions that must hold across all images so they read as a set: colour palette, lighting approach, composition conventions, subject treatment, and level of realism. Write these once as the shared specification.

2. THE IMAGE BRIEF TABLE - one row per image: Slot | What it communicates | Composition | Key elements | Where text sits if any | Aspect ratio

3. THE PROMPTS - one per slot, each built from the shared visual system plus that slot's specifics.

4. THE TEXT OVERLAY PLANNING - if text goes on top, the image needs deliberate empty space with low detail. Specify where, and prompt for it. This is the most common reason a good image fails in use.

5. THE HIERARCHY - which images carry the most weight and should get the most generation effort, and which are secondary and can be simpler.

6. WHAT SHOULD NOT BE GENERATED - slots where a photograph, a diagram, a screenshot, or a chart would serve better than a generated image. Be willing to say that several of these should not be AI images at all. A diagram that explains something is better than an illustration that decorates it.

7. THE GENERATION ORDER - produce the most important image first and establish the look, then generate the rest to match it.

8. THE REJECT RATE - realistically how many generations per usable image for this kind of set, so the effort is not underestimated.

9. THE CONSISTENCY CHECK - after generating, how to assess whether the set holds together: view them together at small size, check colour and lighting consistency, and confirm the subject treatment matches.

10. THE FILE PLAN - naming, dimensions per slot, format and compression for the destination.

Write alt text and captions for these images.

IMAGES: [DESCRIBE EACH, OR PASTE THE PROMPTS THAT MADE THEM]
WHERE THEY APPEAR: [THE PAGE AND ITS SUBJECT]
WHAT EACH IMAGE IS DOING: [informative / decorative / functional / complex data]
PAGE TOPIC AND KEYWORD: [IF SEO MATTERS]

For each image produce:

1. THE PURPOSE CLASSIFICATION first, because it determines everything:
   - INFORMATIVE: conveys content. Needs alt text describing that content.
   - DECORATIVE: adds nothing informational. Should have empty alt (alt="") so screen readers skip it. Describing a decorative image is noise for the user, not helpfulness.
   - FUNCTIONAL: it is a link or a button. Alt text describes the action, not the picture.
   - COMPLEX: a chart or diagram. Needs short alt plus a longer description nearby.

2. THE ALT TEXT - written for someone who cannot see the image but is reading the page:
   - Describe what matters for understanding the page, not everything visible
   - Under about 125 characters where possible
   - No 'image of' or 'photo of'; screen readers already announce it
   - Do not repeat the caption or nearby text verbatim
   - Include text that appears within the image, since it is otherwise lost
   - Natural sentence, ending with a full stop, so screen readers pause correctly

3. THE CAPTION if one is warranted - captions are read by everyone, alt text by some. They do different jobs. A caption adds context, attribution or a point; it should not simply describe the image.

4. THE SEO NOTE - alt text contributes to image search, but it is an accessibility feature first. Include the keyword only where it genuinely describes the image. Keyword-stuffed alt text is both a spam signal and a worse experience for the people it exists to serve.

5. THE FILENAME - descriptive, hyphenated, lowercase.

6. FOR CHARTS AND DIAGRAMS - the short alt plus the full description, which should convey the finding, not just the chart type. 'Bar chart showing sales' is useless; the description should say what the chart shows.

7. THE AI DISCLOSURE - if these are generated images and the context warrants saying so, where that note belongs.

Do not write alt text that describes the image in detail when the page already explains it. Redundancy is a burden for screen reader users.