Building an image prompt
Purpose: turn a wish in plain language into one prompt that separates subject, style and technique, in a single pass.
Based on fokus-KI, "Prompts für KI-Bilder" (April 2026), extended with the limits of the models actually in use here (Z-Image-Turbo, FLUX.2 Klein, SDXL).
When to load
- "Generate an image of …", "write me a prompt for …"
- A result misses and the prompt needs sharpening
- Before the first image, when the request names only a subject
Out of scope: settings and tool choice (the generation skill), storage in a vault (the assets skill), graphics from SVG.
The six building blocks
Order matters: the model weights the beginning more heavily. Subject before style, style before technique.
| # | Block | Question | Example |
|---|---|---|---|
| 1 | Image type | photo, illustration, 3D, painting? | photorealistic, flat vector illustration, 3D render |
| 2 | Main subject | what exactly, with which traits? | tabby cat, green eyes, looking up |
| 3 | Setting | where, with which props? | window sill, blurred garden behind |
| 4 | Time / light | time of day, season, light source | early morning, warm side light from the left |
| 5 | Style | artist, era, mood, palette | in the style of Gustav Klimt, gold tones, calm |
| 6 | Technique | perspective, detail level, composition | close-up, shallow depth of field, rule of thirds |
Rule: 1 and 2 are required, 3–6 are optional. Filling all six when the request does not support it yields an arbitrary image. Four concrete blocks beat six vague ones.
Template
<image type> of <subject with 2–3 traits>, <setting>, <light>,
<style/mood/colors>, <perspective/detail/composition>
Example:
Photorealistic: traditional Japanese tea house beside a still pond, blossoming cherry trees, soft late-afternoon light, warm earth tones, slightly elevated perspective, reflections in the water, calm meditative mood.
One sentence, not a pile of keywords. Punctuation and commas are enough
weighting; brackets, :: and --ar are understood by none of the three
models.
Length and language
- Core 15–50 words. Longer dilutes: attention spreads across subordinate clauses instead of the subject.
- English is the safe default. Style terms (
golden hour,volumetric light,shallow depth of field) and artist names are predominantly English in training data; translations often miss the nuance. - German works with multilingual text encoders for subjects and spatial descriptions. With SDXL or FLUX-class models, translate when in doubt — a German style term can fall flat there.
- Do not mix. A half-German, half-English sentence costs consistency. One language per prompt.
- Open: German versus English at the same seed is not measured yet with these models. Until then English is the default; if a German prompt clearly underperforms, change the language first, not the step count.
Negative prompts: only where the field exists
| Tool family | Negative prompt | Approach |
|---|---|---|
| SDXL | yes | collect wishes there: text, watermark, extra fingers, blurry, deformed |
| Z-Image, FLUX | no | rephrase the negation positively: "no text" → clean surface without lettering |
A "no …" in the positive prompt often works as an invitation: the model reads the noun, not the negation.
Iteration: one variable per run
- First run at low step count, note the seed.
- Name the finding, do not re-describe the image: "steam is missing" is a finding, "doesn't look good" is not.
- Exactly one change — a subject addition, the style, or the steps. Keep the seed, otherwise the comparison is worthless.
- Second run only after a visible flaw. No nudging on suspicion.
- For fine detail (steam, smoke, lettering), raise the step count first, do not lengthen the prompt.
What models cannot do reliably
| Wish | Reality | Way out |
|---|---|---|
| readable text in the image | letters dissolve, especially under six characters | set the text afterwards (SVG, layout tool) |
| hands, fingers, teeth | count and shape are often wrong | crop away the hands |
| exact count ("three candles") | sometimes two, sometimes four | count in the image, otherwise change the seed |
| era ("1750", "Edo") | years are not a feature — the background gains power lines, metal and solar roofs, concrete | set the era via materials: every roof thatched with rice straw or dark cedar shingles, hand-hewn wooden posts, unpaved road |
| brand logo, real face | blurry or distorted | do not aim for it; rights and personality rights |
| "4K", "16:9" in the sentence | read as a style word, not a measurement | set width/height in the tool |
Checklist before sending
- Is the subject within the first ten words?
- Are image type and style named separately (not "a nice picture")?
- At most one style term and one light direction — or deliberately two?
- No negation in the positive prompt?
- Under fifty words, without numbers the tool already sets?
- One language throughout?
Failure modes
| Symptom | Cause | Fix |
|---|---|---|
| subject missing, landscape dominates | setting first or too long | move the subject to the front, cut the setting to five words |
| image feels arbitrary, "stock photo" | only adjectives, no style | name one concrete style or artist |
| everything sharp and flat | no perspective, no light | close-up or slightly elevated plus one light direction |
| colours drift | two style directions fighting | one palette, one style |
| text appears as scribble | text wish inside the prompt | remove it, set the text afterwards |
| result changes wildly despite a small edit | seed or step count changed | fix the seed, hold the steps constant |
| power pole, wires or solar roof in the background | a year does not steer an era; the model fills in infrastructure from training | prescribe materials positively (every roof thatched or dark cedar shingles, hand-hewn wooden posts), not "no power lines" — and check the background, not just the subject |