The Claude art brief: get better AI images

What this actually does

You stop being the prompt writer and start being the art director. Instead of typing "a cozy desk with a plant," you hand Claude one messy sentence and it expands it into a proper brief: subject, camera, lighting, mood, colour palette, and what to avoid. Then you paste that brief into an image model that draws it.

The reason this works: image models reward specificity. A one-line prompt leaves 90% of the decisions to the model, so you get a generic average. A brief makes those decisions for it, so the output matches the picture in your head.

The exact Claude prompt (copy this)

Paste this into Claude, then swap in your own idea on the last line.

You are my art director. I'll give you one rough sentence describing an image I want. Turn it into a detailed image-generation brief with these labelled sections: Subject (who or what, and what they're doing), Composition (camera angle, shot type, framing), Lighting (source, direction, mood of light), Mood (the feeling in one line), Colour palette (3 to 5 specific colours or a named scheme), Style (photo, illustration, 3D render, etc, and any reference era or artist-free descriptor), and Avoid (a short list of things to keep out). Keep it tight and paste-ready, no preamble. Here's my rough idea: [your one messy sentence here]

Claude hands you back a clean brief. Copy the whole thing, paste it into any of the tools below, and generate.

What a good brief actually contains

This is why the prompt above is structured the way it is. Each section removes a decision the model would otherwise guess.

• 🧍 Subject: the specific thing and what it's doing, not just "a woman" but "a woman mid-laugh looking off camera."
• 🎥 Composition: shot type and angle (close-up, wide, low angle, eye level) so the framing isn't random.
• 💡 Lighting: the single biggest lever for a "pro" look. Name the source and direction (soft window light from the left, warm golden hour, hard studio key).
• 🎭 Mood: one line of feeling (calm and editorial, chaotic and fun) that ties the image together.
• 🎨 Colour palette: 3 to 5 named colours or a scheme (warm cream and terracotta). Stops the muddy default look.
• 🖼️ Style: photo vs illustration vs 3D, and any era or texture cue. Ambiguity here is where images go generic.
• 🚫 Avoid: a short exclusion list (no text, no extra hands, no busy background). This alone fixes most "weird AI" artifacts.

The 3 free image tools that actually deliver

All three have a genuinely usable free tier as of July 2026. Each is good at something different, so match the tool to the job.

• ✍️ Ideogram at ideogram.ai. Best for posters, thumbnails, logos and other images where readable text matters. Its official free plan includes 10 slow credits each week, which can produce up to 40 public images depending on the rendering mode. Free generations are public and use the slower queue, so avoid confidential client work (Ideogram pricing, official developer overview).
• ⚡ Leonardo AI at leonardo.ai. Best for volume and control (image references, fine-tuned models, consistent characters). Great when you want to iterate on a look, not just one hero image. Free tier: 150 free tokens per day that reset every 24 hours, roughly 10 to 15 images on a first-party model, and first-party models also have a Relaxed mode that costs no tokens (Leonardo token usage, pricing). Heads-up: tokens don't roll over (use them or lose them daily), free creations are public, and premium third-party models (Flux, Ideogram) burn tokens fast.
• 🔒 Adobe Firefly at firefly.adobe.com. Best for commercially safe images for brand, ads, and client work. Firefly is trained on licensed Adobe Stock and public-domain content, so it's the safest option when the image is going somewhere paid. Free tier: 25 generative credits per month, and 1 credit returns 4 variations, so 25 credits gives up to 100 images to choose from (Adobe generative credits FAQ). Heads-up: full IP indemnification (Adobe covering legal costs if a copyright claim comes up) only applies on paid plans, free-plan images can carry a Content Credentials tag, and credits don't roll over (Adobe Firefly business).

Want variations? Ask Claude to change one variable

Once you have a brief you like, don't restart from scratch. Ask Claude:

Take the brief above and give me 5 versions, changing only ONE variable each time (lighting, then colour palette, then camera angle, then mood, then style). Keep everything else identical so I can compare cleanly.

Now you have 5 controlled experiments instead of 5 random rerolls, so you actually learn what moves the image.

3 bonus prompts to run next

• 🪺 Diagnose a bad image: Here's the prompt I used and here's what came out wrong [describe it]. Rewrite the brief to fix exactly that problem, and tell me which section was the culprit.
• 🎞️ Build a repeatable style: Turn my brand into a reusable image style block I can paste at the top of every prompt: my colours are [x], my vibe is [y], I always want [z] and never want [w]. Give it to me as one tight paragraph.
• 📐 Match a reference without copying it: I'll describe an image I love [describe it]. Write me a brief that captures its lighting, composition, and mood in my own subject, without imitating any specific artist.

How to get the most out of it

One honest caveat