<!-- Free part of the ihumane Graphics and images module, generated from pro/graphics/ by scripts/build-free.js. Edit pro/graphics/, not this file. -->

# 2 · Generate: prompts, sets and the AI look

Image models give you the most average picture of whatever you name. If you
name little, you get the average of everything: a glossy, centred,
over-lit image that looks like a thousand others. A good prompt reads like
a photographer's or illustrator's brief: what's in the frame, where, in
what light, in what medium, and what must not appear.

Use this file only after `1-brief.md` says a generated image is the right
source (B3), and within the image tool's own terms and content rules.

---

### P1 · Prompt with specifics, not adjectives
**Rule:** Write every prompt from the template below. Each line is a
concrete description someone could photograph or draw. Quality adjectives
("beautiful", "stunning", "high quality", "8k", "ultra-detailed",
"masterpiece", "trending") are not descriptions: leave them out.

```
Subject:      who or what, doing what, with which real objects and materials
Purpose:      the job and the slot: "hero image, right half of the page, 4:3"
Composition:  framing, angle, where the subject sits, where the empty space is
              ("medium shot, eye level, subject in the left third, plain wall
              on the right where the headline goes")
Light:        source, direction, time of day, hardness
              ("soft window light from the left, late morning, gentle shadows")
Style:        the medium and its marks: "35mm photo, natural colour, slight grain"
              or "flat vector, 2px even outlines, no gradients". For a set,
              paste the set's style block here word for word (P2)
Palette:      3 to 5 colours from the brand tokens, by name and hex
Avoid:        what must not appear: "no text, no letters, no logos, no
              gradients, no glow, no sparkles, no extra people"
Format:       aspect ratio and size from the brief (1-brief.md B2)
```

If the tool takes a single text field, write the lines as one paragraph in
the same order. If it has a separate negative prompt or "avoid" field, put
the Avoid line there. Use the tool's aspect-ratio setting rather than
asking for a shape in words.
**Why:** Every word you leave out, the model fills with the most common
answer. People recognise those common answers as "AI made this" and
stop believing the page is about something real.
**Agent trap:** a one-line prompt full of praise.

```
Weak:   a beautiful modern image of coffee, high quality, 8k, cinematic
Strong: a ceramic pour-over dripper on a worn oak counter, coffee mid-pour
        into a plain white cup, close-up at counter height, subject in the
        left third, blurred tiled wall on the right, soft morning window
        light from the left, 35mm photo, natural muted colour, slight grain,
        palette of warm oak, off-white and deep brown, no text, no logos,
        no steam swirls, no sparkles, 4:3
```

**Check:** every template line is filled for this image. Delete each quality
adjective; if the result changes nothing you'd describe to a photographer,
it wasn't needed.

> **In ihumane Pro** (https://ihumane.ai/#pricing):
> - P2 · One style block and one recipe for the whole set
> - P3 · Steer away from the generic AI look

### P4 · Never bake important text into a generated image
**Rule:** Ask for images with no text in them ("no text, no letters, no
signage, no labels"). Any words people need to read (headlines, labels,
prices, calls to action, captions, numbers in a diagram) go in HTML or SVG,
placed over or beside the image. Leave a quiet area in the composition for
them (P1, Composition).
**Why:** Text inside pixels can't be read by screen readers, found by
search, translated, copied, resized for a phone or corrected without
regenerating the whole image. Image models also misspell and invent
letters, and one garbled word makes the whole page look careless.
**Agent trap:** prompting for a finished banner with the headline in it.

```
Weak:   a banner that says "Spring sale, 30% off" with flowers
Strong: (image) pale green linen with scattered fresh tulips along the
        bottom edge, large empty area in the top two-thirds, no text,
        no letters
        (HTML) <h2>Spring sale</h2><p>30% off until 30 April</p>, placed
        over the empty area
```

**Check:** zoom into every generated image. Are there any letters, numbers
or letter-like shapes? Is every word on the page selectable text?
**Limit:** share images (`4-place.md` F3) and app icons are pictures of text
by nature. Make them from an HTML or SVG template with real fonts, not from
an image model.

> **In ihumane Pro** (https://ihumane.ai/#pricing):
> - P5 · Inspect every result at full size before you use it