1.2Anatomy of a good prompt
The essentials in 30 seconds
- 7-block template: subject + action + setting + style + light + framing + parameters — each block adds a constraint and reduces randomness.
- Concrete + positive + ordered: describe what must be seen, say what you want (never “no…”, use --no), put the important part first.
- Iterate one block at a time: start simple, add one element, observe — and keep a library of wordings that work.
Running mini-case — Sonia structures her first real prompt
Sonia must illustrate the Hydra café: she starts from “a café” and gets generic randomness. She applies the template: subject (a steaming cup on a wooden table), action (steam rising), setting (a café terrace in Algiers), style (watercolor), light (golden hour), framing (close-up, 85mm). Each addition tightens the result. When a car spoils the scene, she does not say “no car” but adds --no car. She keeps each working version in a reusable prompt notebook.
Sequel to case 1.1 — Sonia moves from the account to the first control lever: structure.
1. The template that steers every dimension of the image
A vague prompt (“a cat”) leaves everything to chance; a structured prompt steers each dimension. Remember the template: [subject] + [action/pose] + [setting/scenery] + [style/medium] + [light/mood] + [framing/lens] + [quality/parameters]. Example: “an old fisherman mending a net, on a harbour quay, documentary photography, soft late-day light, close-up, 85mm lens”.
Sonia applies it to her order: each block she adds (terrace in Algiers, watercolor, golden hour, close-up) reduces randomness. Light and framing terms work better in English (golden hour, rim light, wide shot): better represented in the training data, they are more reliable.
2. The principles that make the difference: concrete, positive, ordered, lean
Be concrete and visual: describe what must be seen (“a sky heavy with dark clouds”) rather than an emotion (“a sad mood”). Prefer the positive: models handle negation badly — “no car” can make a car appear; use the dedicated --no parameter. Order by importance: the first words weigh the most.
Avoid overload: three lines of contradictory elements produce mush — few elements, but precise. Finally, iterate: start from a simple prompt, look at what is missing, add one element at a time. The perfect prompt does not exist: it is a dialogue, and your successes feed your library.
🔎 Concrete example — 🔎 Concrete example — Describe what you see
Concrete, not emotional — like Sonia on her steaming cup.
/imagine a steaming cup of coffee on a wooden table, sunny terrace of a cafe in Algiers, watercolor illustration, golden hour, close-up, 85mm --no car --ar 4:5Reading: each block shows in the image: clear subject, located scenery, warm light, tight framing, intruder excluded.
Lesson: if an addition degrades, remove it — hunt contradictions before stacking.
Section vocabulary
- Prompt structure
- Template: subject + action + setting + style + light + framing + parameters; each block reduces randomness.
Simple example: Sonia's terrace prompt, grown from “a café” to 7 constrained blocks.
What it is not: a vague word that leaves everything to chance.
Reflex: check that no key block is missing before generating.
- Descriptive prompt
- Description of what must be SEEN (concrete visual elements) rather than an abstract emotion.
Simple example: “A sky heavy with dark clouds” preferred over “a sad mood”.
What it is not: a feeling the model cannot see.
Reflex: translate each emotion into a visible sign.
- --no parameter
- Midjourney negation parameter used to exclude an element, more reliable than writing “no…”
Simple example: Sonia's --no car chasing the car off her terrace.
What it is not: writing “without cars” in the prompt text.
Reflex: exclude with --no, describe the rest in the positive.
- Iteration
- Method: start simple, then add one element at a time to measure its real effect.
Simple example: Sonia adds the light, generates, then only the framing.
What it is not: changing everything at once without knowing what helped.
Reflex: one addition = one generation = one observation.
In practice — Hands-on — Build a structured prompt
- First write the subject alone and generate: you get the base and see what the model imagines by default.
- Add the setting and the action, then generate again; then the style/medium; then the light; then the framing.
- At each step, add ONLY one block to measure its real effect on the image.
- Replace any negative wording (“without…”) with a positive description or the --no parameter.
Ready-to-say script — structure a testable prompt (30s)
“I write subject plus action, then setting, then style, then light in English — golden hour — then 85mm framing. To exclude, I add --no car at the end, never no car in the text. I generate at each addition to see what each block changes.”
Say while typing the /imagine block by block; show the effect of a single addition.
Copy-paste template — 7-block /imagine template
/imagine [subject], [action / pose], [setting / scenery in Algiers], [style / medium], [light in English], [framing / lens] --no [element to exclude] --ar 4:5Hydra café cup
Like Sonia: steaming cup + terrace in Algiers + watercolor + golden hour + 85mm close-up + --no car — her first client image.
Fisherman portrait
An old fisherman mending a net, harbour quay, documentary photo, soft late-day light, 85mm close-up: the concrete before the emotion.
English version
Same idea translated with technical terms in English (rim light, shallow depth of field, wide shot): compare and keep the version that obeys best.
Anti-overload
Confused prompt with 20 elements: cut it to 5, hunt minimalist vs rich in details, then rebuild one block at a time.
✓ Best practices
- Describe what must be seen (concrete, visual) and put light/framing technical terms in English.
- Put the important elements at the start of the prompt: the first words weigh the most.
- Iterate one block at a time and keep your successful prompts in a reusable library.
✗ Common pitfalls
- Writing negations as text (“no…”, “without…”) instead of using --no.
- Overloading the prompt with contradictory elements that produce a blurry compromise.
- Describing an abstract emotion instead of observable visual elements.
Key takeaways
- Template: subject + action + setting + style/medium + light + framing/lens + parameters.
- Describe what must be SEEN (concrete, visual), not an abstract emotion; the first words weigh the most.
- Avoid negations in the text (use --no) and overload with contradictory elements.
- Iterate by adding one element at a time; keep a library of prompts that work.
48h Challenge — apply tomorrow
Take one real client subject: generate it alone, then add 4 blocks one by one (setting, style, light, framing) with the template above. Replace one negation with --no and archive the winning version.
- I build every prompt with the 7-block template.
- I write in the positive and I exclude with --no, important part first.
- I iterate one block at a time and I archive my winning wordings.
Frequently asked questions
Should I write my prompts in English?
It is no longer essential, but English often stays more reliable for technical vocabulary (golden hour, rim light, bokeh, wide shot), massively present in the training data. Hybrid recommendation: base description in your language, light/framing/style terms in English. If a result disappoints, translate and relaunch: the gap is sometimes spectacular.
Why does my very detailed prompt give a confused result?
Through dilution (20 elements = each one approximate), implicit contradictions (minimalist + rich in details), saturation (the end of the prompt weighs less) and object mixing. Fix: restart from a short prompt that works, add back one by one, and for complex scenes generate a base then retouch (4.1) or compose (4.5).
How do you handle a negation in a Midjourney prompt?