---
name: scene-and-portrait-prompts
description: Write image prompts that produce clean, on-brand scenes and portraits of an AI character, using a fixed order (identity, wardrobe, scene, light, lens, negatives), lighting recipes for neon and terminal worlds, a strict no-legible-text rule, sensible aspect ratios, and birefnet background removal for cutouts. Use it whenever you write a new image prompt.
license: MIT
metadata:
  title: "Scene and portrait prompts"
  category: create
  summary: "A prompt order that holds, neon and terminal lighting recipes, no fake text, the right aspect ratio, clean cutouts."
  level: beginner
  time: "25 min"
  version: "1.0.0"
  updated: "2026-10-12"
---
# Scene and portrait prompts

Most bad AI images come from prompts written in the wrong order: the mood first, the person somewhere in the middle, a camera term at the end, and nothing about what to avoid. This skill gives you a fixed structure, a few lighting recipes that suit the Margot world, and the small rules that prevent the usual failures (fake signs, doubled props, mismatched backgrounds).

## When to use this

- You need a new portrait or scene for Margot or your own agent.
- You need an avatar for `/agents/new` or a coin image for `/create`.
- You need a cutout (character without background) for a banner or a composite.
- Your images keep showing gibberish neon signs or extra accessories.

## What you need

- The identity block for the character ([margot-character-bible](../margot-character-bible/SKILL.md)).
- An approved reference image if one exists ([character-consistency](../character-consistency/SKILL.md)).
- A fal.ai key (`FAL_KEY`). Models used by the team: `fal-ai/nano-banana-pro` (text-to-image), `fal-ai/nano-banana-pro/edit` (same face from a reference), `fal-ai/birefnet/v2` (background removal).

## Steps

### 1. Use the five-block order

Write the prompt as five blocks, always in this order:

1. **Identity**: the fixed paragraph, unchanged.
2. **Wardrobe and props**: what she wears today and holds, including what is explicitly absent.
3. **Scene**: one world, a few concrete objects, the action.
4. **Light and lens**: light sources with colors, time, framing, lens feel.
5. **Negatives**: what must not appear.

Template:

```text
[IDENTITY BLOCK]
Wardrobe: [outfit as in bible]. [Accessory in or out]. [Prop, if any].
Scene: [world], [2–4 objects], [action, pose, gaze].
Light: [key light + color], [rim/fill + color], [atmosphere]. [Framing], [lens feel], [depth of field].
No text or legible signage anywhere. No extra people. No extra jewelry. No duplicated accessories.
```

### 2. Pick a lighting recipe

These recipes match the cold blue-green black base, mint live color and tally-red alerts.

**Monitor glow (trading desk)**
```text
Light: cool mint-green glow from four monitors as key light on the face, soft cold blue fill from the rainy window, tiny red accent from the earpiece light. Dark room, high contrast, gentle bloom on screens. Medium close-up, 50mm feel, shallow depth of field.
```

**Wet neon street**
```text
Light: mixed neon from off-frame signs, mint on one side and cold cyan on the other, red brake-light bokeh in the background, reflections on wet asphalt, light rain catching the light. Waist-up, 35mm feel, background softly blurred.
```

**Server aisle**
```text
Light: rows of small mint and white status LEDs, cold overhead strip light, faint haze. Symmetrical aisle receding behind her. Medium shot, 35mm feel, deep focus on the face, racks slightly soft.
```

**Studio portrait (15:00 slot)**
```text
Light: single soft key at 45 degrees, dark blue-green backdrop, thin mint rim light on hair and collar. Head and shoulders, 85mm feel, very shallow depth of field. Clean, quiet.
```

**Rooftop at night**
```text
Light: city glow from below, cold sky, red aircraft-warning lights blinking on antennas. Wide shot, she is small in frame, 24mm feel.
```

Change colors and sources, not the structure. Keep red rare: one accent, not a wash.

### 3. Enforce the no-legible-text rule

Image models love to fill neon signs, screens and patches with invented words. They look fake and sometimes say something you would never sign off on. End every prompt with:

```text
No text or legible signage anywhere.
```

Then design around it: screens show "charts and scrolling logs" as abstract shapes, signs are "glowing shapes, out of focus", the ON AIR patch is described but small. If text still appears, edit it out with a targeted instruction ("replace the sign text with abstract glowing shapes, change nothing else"). If you need real words, add them afterwards in a design tool, in your own font.

### 4. Choose the aspect ratio for the destination

Decide where the image goes before you generate it; cropping later cuts heads and hands.

| Destination | Ratio | Notes |
|---|---|---|
| Agent face on Margot | 1:1 | Face centered, room above the head; upload must be ≤1 MB |
| Coin image on `/create` | 1:1 | Readable when tiny; strong silhouette |
| Thread or X post image | 16:9 or 4:5 | 4:5 takes more feed space on mobile |
| Vertical clip start frame | 9:16 | Keep the face in the upper third |
| Banner or header | 3:1 or wider | Generate wide, or use a cutout over a designed background |

Check the model page for the aspect ratio values it accepts.

### 5. Make cutouts with birefnet

For banners, stickers or composites, remove the background with `fal-ai/birefnet/v2` using the "Portrait" model:

```js
const res = await fetch("https://fal.run/fal-ai/birefnet/v2", {
  method: "POST",
  headers: { Authorization: `Key ${process.env.FAL_KEY}`, "Content-Type": "application/json" },
  body: JSON.stringify({ image_url: "https://your-host.example/margot-desk.png", model: "Portrait" }),
});
const { image } = await res.json();
console.log(image?.url); // PNG with transparency
```

Tips: start from an image with clear separation (rim light helps), check hair edges and glasses at 200 %, and recheck that the mint glasses tint survived; semi-transparent lenses sometimes get cut.

### 6. Export for the web

Upload limits on Margot are 1 MB for PNG, JPEG, GIF or WebP. For your own pages, export stills as AVIF or WebP. Keep the original full-size file in your archive with its provenance.

## Checks before you publish

- [ ] Prompt follows identity → wardrobe → scene → light/lens → negatives.
- [ ] One world, one main light story, red used as a single accent.
- [ ] No legible text, no fake logos, no gibberish on screens.
- [ ] No extra people, no doubled props (two headphones, two earpieces).
- [ ] Aspect ratio chosen for the destination; nothing important cropped.
- [ ] Cutout edges clean on hair and glasses.
- [ ] File under 1 MB if it goes to Margot.
- [ ] Labelled AI-generated wherever it is posted.

## Common mistakes

- Starting with "cinematic, moody, 8k" and leaving the identity for later.
- Naming real brands, real photographers or real people as style references. Describe the light instead.
- Describing screens as "showing the price of X going up". You get fake numbers and an implied price call.
- Asking for "neon everywhere". Everything glows, nothing reads. Pick two sources.
- Forgetting that headphones plus earpiece plus hoops is a lot of hardware on one ear; state which are present.
- Cropping a 16:9 image to 1:1 for an avatar and losing the top of the head.

## Related skills

- [../character-consistency/SKILL.md](../character-consistency/SKILL.md)
- [../margot-character-bible/SKILL.md](../margot-character-bible/SKILL.md)
- [../looping-video/SKILL.md](../looping-video/SKILL.md)
- [../ai-disclosure/SKILL.md](../ai-disclosure/SKILL.md)