Illustration Prompt Generator

If it's for a website, ask which page and what surrounds the image (dark/light theme, what content is above/below). If you already know the project context from the conversation, reference it directly instead of asking generic questions.

Step 2: Style

If style is already established from earlier in the conversation, confirm it briefly ("Same voxel art with high detail?") and move on.

Otherwise, present options and let the user pick or describe their own:

• Voxel art (cubic, Minecraft-like, MagicaVoxel renders)
• Isometric illustration (angled top-down view, architectural)
• Flat / vector illustration (clean shapes, minimal shadows)
• Watercolor / hand-painted (organic, textured)
• Line art / ink drawing (outlines, hatching)
• 3D render (realistic materials, studio lighting)
• Photorealism (looks like a photograph)
• Collage / mixed media
• Pixel art (retro, 8-bit / 16-bit)
• Other — user describes

Ask about detail level: minimal/clean vs. highly detailed/rich.

Step 3: References (recommended, not required)

References dramatically improve prompt accuracy — but they're optional.

If this is a continuation — you already have references and style context from earlier in the conversation (or from a previous illustration in the same project) — summarize what you already know and ask: "I already have your references and style preferences from before. Want to show me new references, or should we skip this step and use what we have?"

If the user says skip, move on. If they add new references, incorporate them.

If this is the first image, tell the user:

"Reference images help me write a much more precise prompt. If you have examples of a visual style you like, share 3 or more — and for each one, describe what specifically you like (color palette, detail level, composition, mood, texture, lighting). The more specific you are about WHY you like each reference, the better the prompt.

If you don't have references right now, we can skip this step — but without them, the generator decides style details on its own, and the result may not match what you had in mind."

If the user provides references, analyze them and summarize back:

• Common patterns across the references (shared color temperature, detail density, perspective, etc.)
• Key differences between them
• Which elements you plan to emphasize in the prompt

Confirm your interpretation before moving on.

If the user skips, acknowledge it and move on. Compensate by being extra specific in Steps 4 and 5 — ask more detailed questions about composition and mood to make up for the missing visual anchors.

Step 4: What to Depict

Based on the context (Step 1) and style (Step 2 + Step 3 references), propose 2-3 specific scene ideas. Describe each in 1-2 sentences so the user can visualize the options.

The user picks one, modifies, or describes something entirely different. Either way, nail down the specific objects, their arrangement, and the story the image tells.

Step 5: Colors and Mood

If palette and mood are already established from a previous illustration in the same project, confirm briefly ("Same dark theme #0D0D0D, accent #52C492, vivid character colors?") and move on. Only ask if something might differ for this specific image.

Otherwise, ask about:

• Theme: dark / light / neutral background
• Specific colors: if the project has brand colors, ask for HEX codes
• Mood: premium, cozy, energetic, playful, mysterious, clinical, warm, cold
• Lighting: natural, studio, golden hour, neon, moody, dramatic shadows, flat

If you know the project's design system (CSS variables, brand guidelines), reference those colors directly.

Step 6: Format and Size

If format is already established (e.g., same panoramic banner as previous illustrations), confirm briefly and move on.

Otherwise, ask about:

• Proportions: 16:9 standard, 3:1 panoramic banner, 1:1 square, 4:5 vertical, 21:9 ultrawide, custom
• Resolution: suggest appropriate size based on context (e.g., 2560x800 for a wide banner, 1920x1080 for standard, 1080x1080 for social)
• Background: transparent PNG, solid color (which HEX), or scene fills the frame

Output: File Name + Prompt

After all 6 steps, first generate a file name, then the structured prompt.

File name

Suggest a short, descriptive file name in kebab-case (lowercase, hyphens, no spaces). The name should make it immediately obvious what the image is and where it goes.

Format: hero-people.png, card-resp-007.png, bg-insights-barriers.png, cover-presentation-q1.png

Pattern: [placement]-[subject].png

Present the name before the prompt so the user knows what to call the file.

Prompt structure

Generate a single structured prompt in English. The prompt has 8 blocks:

SCENE: [Overall composition — what the viewer sees, camera angle, how elements are arranged across the frame]

STYLE: [Art style, detail level, rendering approach, reference anchors like "MagicaVoxel render" or "Studio Ghibli watercolor"]

OBJECTS: [Every specific object in the scene — what it is, where it sits, approximate size relative to others, key visual features]

COLORS: [Exact palette with HEX codes where possible, how colors distribute across the scene, highlight/accent colors, shadow tones]

ATMOSPHERE: [Lighting direction and quality, mood, depth/fog, ambient effects like glow or particles]

EDGES: [How the scene terminates — fades to background color, hard crop, vignette, floats as an island, bleeds off-frame]

FORMAT: [Width x Height in pixels, aspect ratio, background type (transparent / solid #HEX / scene-filled)]

NEGATIVE: [What must NOT appear — watermarks, UI elements, and scene-specific unwanted elements. Text defaults and anatomy defaults are added automatically per the rules below]

Vehicle defaults

When any vehicle (car, truck, bus, van) appears in the scene, always apply these defaults in both OBJECTS and NEGATIVE blocks unless the user explicitly states otherwise (e.g., "door is open", "door removed", "cutaway view"):

• Windows: transparent. The interior and occupants must be visible through the glass.
• Doors: closed, solid panels. Generators tend to omit doors or render cutaway views — counteract this by explicitly stating "doors fully closed, solid door panels" in the OBJECTS block and "no cutaway vehicles, no missing doors, no exposed interior through door openings" in the NEGATIVE block.
• Proportions: balanced front-to-back. Explicitly describe that the hood and trunk are proportional, the car is not stretched or compressed. If the car model matters, name it or describe its silhouette (sedan, SUV, hatchback).

These rules exist because AI generators consistently produce cars without doors or with cutaway interiors when not explicitly instructed. Treat this as a hard default — only override when the user's scene description requires it.

Human figure defaults

When any person (full body, portrait, or partial figure) appears in the scene, always apply these defaults in both OBJECTS and NEGATIVE blocks unless the user explicitly states otherwise:

• Hands: five fingers each, anatomically correct. This is the single most common AI generator failure. Always explicitly state "five fingers on each hand" in the OBJECTS block. If hands are visible, describe their exact pose: "right hand gripping a mug handle, left hand resting flat on the table." The more specific the pose, the fewer deformations.
• Hide hands when possible. If hands are not essential to the scene, place them in pockets, behind an object, or out of frame. This avoids the problem entirely.
• Joints bend correctly. Explicitly state that elbows and knees bend in natural directions. In the NEGATIVE block: "no backwards-bending joints, no extra limbs."
• NEGATIVE block must include: "extra fingers, fused fingers, missing fingers, deformed hands, six fingers, extra limbs, extra arms, missing ears, deformed anatomy."

Full-body shot defaults

When the prompt requires a full-body view of a person or character, apply these additional checks on top of the Human figure defaults:

• Feet or footwear must be visible and described. Generators frequently crop at the ankles. Explicitly state footwear type and the surface they stand on: "brown leather boots on a cobblestone path."
• Hair and head fully in frame. Describe hairstyle or headwear to anchor the top of the figure.
• Ground or floor surface described. The figure needs a surface to stand on — name the material, color, and condition: "cracked concrete floor", "wet grass."
• Suggest vertical or tall aspect ratio. Full-body shots need vertical space. Recommend 2:3, 3:4, or 9:16 unless the scene context requires otherwise (e.g., a figure in a wide landscape).

Text defaults

AI generators (as of March 2026) cannot reliably render text. Even two-word signs come out misspelled or garbled.

• If the user did NOT request text in the image: always add to the NEGATIVE block: "no text, no letters, no words, no numbers, no labels, no signs with writing, no watermarks." This prevents random text artifacts from appearing on surfaces, signs, or clothing.
• If the user DID request text: warn them before generating. Say: "Text rendering is unreliable in current AI generators — most models misspell even short words. I recommend generating the image without text and adding the text in post-production (Figma, Canva, Photoshop). If you still want text in the prompt, I'll include it, but expect to need several attempts." If they proceed, put the exact text in quotes in the OBJECTS block and keep it to 1-2 short words maximum.

Counting defaults

When the scene includes 3 or more identical objects (chairs, trees, windows, people in a crowd), generators frequently produce the wrong number — adding extras or merging objects together.

• Describe each object individually with its position. Instead of "four chairs around a table," write: "Chair 1: left side of the table, facing right. Chair 2: right side of the table, facing left. Chair 3: far side of the table, facing the viewer. Chair 4: near side of the table, with its back to the viewer."
• State the exact count in both OBJECTS and SCENE blocks. Redundancy helps: "exactly four chairs, no more, no fewer."
• NEGATIVE block: "extra [object name], duplicated [object name], merged [object name]."
• For large numbers (5+): warn the user that precise counts become unreliable. Suggest using words like "several" or "a group of" if exact count is not critical, or suggest compositing multiple generations.

Prompt quality rules

• Minimum 150 words. Below this threshold, generators fill in too many details on their own, producing inconsistent results. There is no upper limit — write as much as the scene requires.
• Be concrete, not abstract. "A tall building" is weak. "A 12-story navy blue (#2A5A8C) office tower with visible voxel-block construction, darker left face, lighter top face" is strong.
• Use HEX colors whenever the user provided them or when the project has a known palette. Generators increasingly respect color codes.
• Name the perspective explicitly. "Isometric 30-degree tilt", "bird's eye view", "eye-level straight on", "low angle looking up."
• Describe light direction. "Lit from upper-left" or "backlit with rim lighting" — never leave lighting to chance.
• Include scale anchors. If the scene has multiple objects, describe relative sizes: "the central tower is 3x taller than the surrounding houses."
• Reference patterns from the user's reference images. If the user loved the "rich detail density" of a reference, explicitly call for high detail density in the prompt. Connect the prompt back to what the user showed you.

After delivering the prompt

Ask the user to generate and share the result. If they want adjustments, iterate — modify specific blocks of the prompt rather than rewriting everything. This teaches the user which blocks control which aspects of the output.