image-prompting

Files (2)

• SKILL.md
• prompt-templates.md

SKILL.md (first lines)

# Image Prompting — Playbook for Generation Prompts

A reusable method for turning a content brief into a precise, repeatable prompt for a modern instruction-following image generator (the gpt-image family and equivalents). Load this whenever a task needs a visual asset: a blog cover, a video thumbnail, a social-post image, an inline illustration, a diagram/infographic, a logo mark, or an edit of an existing image.

This skill produces a **prompt and a verification pass**, not the image itself. It does not call any rendering API — it gives the caller the text to feed one, plus the size/quality settings to request and a checklist to judge the result.

## Core Principle

A generator renders what it can resolve from the words it is given. Two failure modes dominate: **under-specification** (it invents the parts you left blank) and **drift** (across edits or iterations, the parts you wanted kept silently change). The whole playbook is built to defeat both:

1. **Name every load-bearing decision.** Subject, medium, framing, light, palette, and any literal text are decisions — if you do not state them, the model decides for you, differently each run.
2. **Separate what must change from what must stay.** State invariants explicitly and repeat them on every iteration. This is the single highest-leverage habit; skipping it is why "just make the sky bluer" also moves the horizon.
3. **Change one variable per pass.** Isolated edits are debuggable; rewrites are not.

## The Loop

```
intake → spec → prompt → render → verify → (refine one variable) → ship
```

Stay in `spec` until the brief is unambiguous. Most wasted renders trace to a thin spec, not a weak model.

---

## 1. Intake — brief to spec

Before writing a single prompt word, resolve these from the task brief (ask only if the answer changes the image; otherwise pick a sane default and note it):

| Slot | Question | Default if silent |
|------|----------|-------------------|
| **Purpose** | Cover, thumbnail, post image, illustration, infographic, logo, edit? | infer from where the asset will be used |
| **Subject** | What is literally in the frame? | the brief's headline noun |
| **Placement** | Where will it live (platform, page region)? | drives aspect ratio — see §10 |
| **Text-in-image** | Any words rendered *inside* the image? Exact string? | none — keep text out, overlay later |
| **Brand/style anchor** | A palette, a prior asset, a house style to match? | neutral, clean, modern |
| **Mood** | One or two adjectives for the feeling | matches the content tone |
| **Hard noes** | Anything that must NOT appear | no watermark, no stray text, no logos |

State the **purpose explicitly in the prompt** ("a blog cover image", "a UI mockup", "an infographic") — it switches the model into the right rendering mode and changes its defaults for layout, text density, and realism.

---

## 2. Prompt Anatomy

Order the prompt from the outside in, so the model establishes the scene before it places detail. A reliable slot order:

```
[purpose/medium] → [scene/background] → [subject + pose/action] →
[key details] → [composition/framing] → [light] → [palette/mood] →
[literal text, in quotes] → [negative constraints / invariants]
```

Format is flexible — a clean descriptive paragraph, a line-per-slot list, or a tagged/structured block all work. For anything you will reuse or hand to another agent, **prefer the line-per-slot form**: it is maintainable, diffable, and easy to vary one slot at a time. Cleverness in syntax buys nothing; clarity does.

A minimal prompt is often enough; add detail only where the default would be wrong. Long prompts work but are harder to debug — grow them by isolated additions, not by front-loading every adjective you can think of.

---

## 3. Composition

Composition is the cheapest lever on perceived quality. Specify:

- **Framing / crop** — close-up, medium, wide, overhead/flat-lay, extreme close-up macro.
- **Camera angle** — eye-level, low angle (subject looms), high angle (subject diminished), Dutch tilt for tension.
- **Subject placement** — "centered", "rule-of-thirds, subject left", "subject lower-right with negative space upper-left for overlay text".
- **Depth** — clear foreground / midground / background, or flat with no depth for graphic styles.
- **Negative space** — reserve it deliberately when text or a logo will be composited on top later. This is the move that makes covers and thumbnails usable.

For people and creatures, also pin **body crop** ("full body, feet included" vs "head and shoulders"), **gaze** ("looking off-frame, not at camera"), **scale** ("child-sized against the table"), and **interaction** ("hands gripping the handle"). These fix the three things generators get wrong most: proportion, action geometry, and where the eyes point.

---

## 4. Style & Medium

State the **visual medium** first — it is the largest stylistic switch:

- Photograph, 3D render, flat vector, watercolor, oil, gouache, ink line art, pixel art, matte painting, isometric, paper-cut/collage, claymation.

Then layer **style qualifiers** only as needed: brushstroke texture, film grain, halftone, cel shading, soft outlines, high detail vs minimal. Add quality levers (grain, macro detail, textured strokes) sparingly and only when the bare medium reads too clean.

Name a **concrete reference register** rather than an artist's name where possible ("editorial magazine illustration", "children's picture-book", "technical blueprint", "premium product photography") — it transfers a whole aesthetic without copying any one person and keeps the output reusable.

For **character/brand consistency across a series**: lock an anchor description once (appearance, proportions, outfit, palette, demeanor) and on every later image instruct "same character, do not redesign — new scene only", restating the anchor traits each time. Style continuity must be named, not assumed.

---

## 5. Camera & Lens (for photoreal)

Reach for these only when the target is photographic realism:

- Use the explicit word **"photorealistic"** to engage the realism pathway.
- **Focal length** — wide (24–35mm) for environments and a sense of place; standard (50mm) for natural framing; portrait/tele (85–135mm) for flattering subjects and compression.
- **Aperture / depth of field** — "shallow depth of field, background bokeh" vs "deep focus, everything sharp".
- **Film / sensor cues** — "35mm film", "film grain", "natural color balance".
- **Authenticity markers** — pores, skin texture, fabric wear, weathering, slight asymmetry. Their absence is what makes an image read as plastic/AI.
- **Avoid** studio gloss and over-retouching unless the brief is explicitly a polished product shot.

---

## 6. Light

Light sets realism and mood more than any other single factor:

- **Direction** — front, side (sculpts form), back (rim/silhouette), top, three-quarter key.
- **Quality** — soft/diffused (overcast, softbox) vs hard (direct sun, single bulb) with crisp shadows.
- **Time / temperature** — golden hour (warm, long shadows), blue hour, midday (flat, hard), neon night, candlelit.
- **Special** — volumetric god-rays, rim light, practical light sources in frame, high-key (bright, airy) vs low-key (dark, dramatic).

For edits that change weather or time of day, change **only** the light and atmosphere and explicitly preserve camera angle, object positions, and scene identity.

---

## 7. Mood & Palette

- **Mood** — one or two adjectives carried through the whole prompt: serene, tense, playful, austere, nostalgic, energetic. Let it steer light, color, and pose rather than stating it once and contradicting it elsewhere.
- **Palette** — name it concretely: "muted earth tones", "high-contrast black and red", "pastel triad", "monochrome teal", or give explicit hex anchors when matching a brand. State whether it is warm or cool, saturated or desaturated, high or low contrast.
- For **atmospheric wide scenes** (cinematic, dark, rainy, neon), amplify scale, atmosphere, and color to hold the mood even when surface realism softens.

---

From Arcanada-one/datarim, MIT.

More from Arcanada-one/datarim