Lesson 2 of 9 · about 15 minutes
How Image Prompts Steer the Picture
You'll be able to
- Name the six working parts of an image prompt and what each controls
- Predict how a vague prompt and a specific one will differ
- Understand why factual text belongs in an editor, not in the model's output
- Understand how 1024x1024, 1080x1920 and 1920x1080 frames change composition and safe areas
An image model behaves like an illustrator who is fast, tireless and has never been to your town. It draws exactly as much as it is told and guesses the rest. Knowing what each part of a prompt controls explains why one AI image looks considered and another looks like stock art from nowhere, and it helps when judging images other people put in front of you.
The six working parts
| Part | What it answers | Example words |
|---|---|---|
| Subject | Who or what, doing what | an older woman reading a water bill at a kitchen table |
| Setting | Where and when | small farmhouse kitchen, winter evening |
| Style | What kind of picture | flat vector illustration, gouache painting, paper cutout, documentary photo |
| Light | Mood and time | single warm overhead bulb, low golden sun, overcast and soft |
| Composition | Framing and camera | wide shot, subject on the left third, eye level, empty wall on the right |
| Text and extras | Lettering, palette, things to avoid | muted blue and tan palette; no text |
Order matters a little. Most models weight early words more, so the subject first and fussy details last tends to work. One sentence or a short comma list both work. Plain description, the way a caption reads, gets better results than instructions or pleading.
Vague versus specific
Example: a road bond explainer
Vague: road bond, potholes, cool image, ultra realistic, award winning
What comes back: a random photoreal street, maybe a city nobody local has ever seen, maybe a giant pothole with lens flare. Photoreal also means it can be mistaken for a real road in a real town, which is a problem for anything about a public vote.
Specific: "Flat vector illustration of a two-lane county road with three patched potholes, a yellow road grader parked on the shoulder, rolling dry hills behind, late afternoon light with long shadows, wide shot with the road running from lower left to upper right, open pale sky across the top third for a headline, muted ochre and slate palette, no text."
The parts: subject (road, potholes, grader), setting (hills), style (flat vector), light (late afternoon), composition (diagonal road, sky for headline), extras (palette, no text). Six for six.
The specific version is 59 words. In practice that is near the upper end of useful. No vendor publishes a hard cutoff, but much longer prompts tend to lose details as words compete for the model's attention.
Style
"Watercolor photo cyberpunk oil painting" gives mush, because the model averages competing styles. One named medium works far better. For public-issue material, illustration styles have a built-in advantage: nobody mistakes a paper cutout for a news photo. Realism for a scenario ("what a dry reservoir could look like") needs a heavy label, covered in lesson 4.
Prompting with the names of living artists is a weak shortcut. Many tools block it, and it borrows a working artist's identity. Describing qualities ("thick brush strokes, limited palette, visible paper texture") gets a similar effect without the problem.
Light does the emotional work
Change only the light and the same scene tells a different story.
- "Bright midday sun, hard shadows": plain, factual, a little harsh.
- "Low winter sun through a window, long blue shadows": quiet, lonely.
- "Single bare bulb overhead, dark corners": tense, late-night.
- "Soft overcast light": neutral, good for explainers that should not editorialize.
For neutral civic explainers, soft overcast or even studio light is the honest default. Dramatic light editorializes, and viewers feel it even when they cannot name it. When an image about a contested issue is lit like a thriller, the lighting is part of the argument.
Composition and the three ImageStudio frames
ImageStudio has a field to "Add a composition hint for this shape" and a "Show platform safe areas" toggle on the result. Both exist because the frame shape decides where the subject and any text can sit.
- 1024x1024 (square). Center works here. A subject sitting a bit high leaves room for a caption bar below. Typical prompt words: "centered, medium shot, simple background."
- 1080x1920 (vertical). Things stack: headline space at the top, subject in the middle, nothing important in the bottom fifth. Typical prompt words: "tall vertical composition, subject in the middle of the frame, open sky at the top, plain ground at the bottom."
- 1920x1080 (horizontal). The subject sits on the left or right third with calm space on the other side. Typical prompt words: "wide shot, subject on the left third, empty wall on the right."
Worked example: where the safe zone falls
Say a reels app covers roughly the bottom 20 percent of a vertical video with captions and the account name. On a 1920-pixel-tall frame, 1920 x 0.20 = 384 pixels. Anything important needs to sit above row 1536. Exact zones differ by app and change often, which is why the safe area overlay is more reliable than a number from a lesson. The arithmetic shows how much of the frame is effectively gone.
Text in the image
The 2026 leaders, OpenAI's Images 2.0 among them, render short text well. OpenAI's launch page specifically touts better text rendering. Short means a headline of a few words, a shop sign, a label on a jar.
Where text still goes wrong:
- Exact wording works best in quotes:
a hand-painted sign that reads "OPEN HOUSE SAT". Without quotes the model paraphrases. - Past about six words, every extra word is another chance for a misspelling.
- Letters that look right in a thumbnail turn out, at full size, to be an "R" with an extra leg.
- Facts (a date, a dollar figure, a name, a URL) are guesses when a model renders them. Careful users generate the image with empty space and set the type in an editor, because nobody should let a model guess at a meeting date.
Negative prompts where supported
A negative prompt lists what to keep out. Support varies:
| Tool | How things get excluded |
|---|---|
| Midjourney | --no followed by words, for example --no text, people. Each word is read separately, so "--no modern clothing" can be read as "no modern" and "no clothing." |
| Stable Diffusion and many open-weight setups | A separate negative prompt field in most interfaces |
| Chat-style generators (ChatGPT, Gemini) | No separate field. Exclusions go in the instruction itself: "Do not include any text or logos." |
| ImageStudio | The page shows one Prompt field and no negative field, so exclusions go in the prompt, kept short |
Positive description usually beats any negative. "Empty street at dawn" beats "street, no people, no cars," because it gives the model something to draw instead of something to avoid.
What a usable image prompt contains
- The subject first, in plain words
- A setting with place type and time of day
- One named style
- Light chosen on purpose
- Composition written for the actual frame size
- Text kept short and quoted, or left out for an editor
- No living artist names, no hype words
Reporting angle
A prompt is a record of intent. When a public body or campaign publishes an AI image, ask for the prompt, the tool and the settings, and compare them to what the image shows. Check any text inside the image against a source, since generated lettering and numbers are often wrong. Ask whether the prompt named a living artist or a real person, which raises licensing and likeness questions.
Research angle
- Midjourney docs: No: how negative prompting works.
- OpenAI API docs: Image generation: options, edits and limits for the image API.
- ImageStudio: the Journal Tools image generator, which strips metadata from its output.
Check yourself
1. Which part of a prompt usually does the most to change the mood of an image without changing what is in it?
Light sets mood. Same porch, same chair: noon sun reads plain, low winter sun reads lonely, a single bulb reads tense.
2. Why do careful users leave empty space in an AI image and set dates, dollar figures and names in an editor afterward?
Short text renders well now, but every extra word is another chance for a misspelling, and a model has no idea what the real meeting date is. Type set by a person keeps spelling and facts under control.
3. Midjourney's docs recommend which way of keeping fruit out of a still life?
Writing 'without fruit' puts the word fruit in the prompt. Midjourney's --no parameter is the supported way to steer away from it.
4. If an app covers roughly the bottom 20 percent of a 1920-pixel-tall vertical frame, how many rows are lost?
1920 x 0.20 = 384 rows, so anything important has to sit above row 1536. Exact zones differ by app and change often.
5. Which prompt gives a model the most to act on for an illustration of a library board meeting?
It names subject, setting, style, light, composition and palette. Hype words like 'amazing' and '8k' add nothing a model can act on.
6. Why does an obviously stylized illustration (flat vector, watercolor, paper cutout) suit material about a public issue?
A clearly drawn look is its own first label. Photoreal illustration invites people to believe they are seeing a real scene.
Terms in this lesson
- Prompt
- The text instruction given to an image model describing what to make.
- Negative prompt
- A separate instruction naming things to leave out. Only some tools support it, through a field or a parameter like Midjourney's --no.
- Rule of thirds
- A framing guide that splits the frame into a 3 by 3 grid and places the subject on a line or crossing instead of dead center.
- Negative space
- Calm, empty area in a picture. In design, the place where text or a logo can sit.
Sources
- Midjourney docs: No (negative prompting) checked 2026-10
- OpenAI: Introducing ChatGPT Images 2.0 (text rendering) checked 2026-10
- ImageStudio (Journal Tools) checked 2026-10