The Claude creative director prompt

Your images are not failing at the model

When an AI image looks fake, the usual reaction is to switch tools. It is almost never the tool. Every current model can produce something genuinely good, and the same model will produce something flat and plastic five minutes later, from the same person, in the same session. The variable that changed is the brief.

There is a layer between "I want a photo of a cocktail on a bar" and a finished image, and in a real production that layer has a name and a person attached to it. The creative director decides the mood, the light, and where the eye lands, and hands the photographer a brief. Nobody teaches that layer to people using AI tools, so everyone skips straight from idea to prompt and wonders why the output looks like stock.

The fix is not learning secret prompt words. It is putting the direction layer back in, and letting Claude write it, because writing a shot brief is a language job and language is what Claude is for.

The 3 layers a good brief always has

1. 🎭 The vibe. Mood, setting, and the story behind the shot. Not adjectives stacked up, an actual situation. "A cocktail on a bar" is a subject. "The last drink of a slow Tuesday evening in a small neighbourhood bar, nobody else in the room yet" is a scene, and a scene gives the model something to be consistent about.
2. 💡 The lighting intention. Where light comes from, what quality it has, and what feeling that creates. This is the single biggest gap between an image that reads as photographed and one that reads as generated, because generated images default to flat, even, sourceless light that exists nowhere in real life.
3. 🎯 The subject direction. Angle, distance, where the subject sits in the frame, and where the empty space goes. This is the part that decides whether the image works as a post, a thumbnail, or a background for text, and it is the part people leave entirely to chance.

The creative director prompt

This is the one I paste at the start of a session. It makes Claude interview you first, which is the whole point, because a director who does not ask questions is just guessing on your behalf.

You are my creative director for an image I am about to generate. You do not generate the image. You write the brief that a generation model will execute, and your job is to make decisions I have not thought about yet.

Here is what I want: [describe the image in one plain sentence, however rough]. It will be used for [where it goes: an Instagram post, a website hero, a thumbnail, a product page] and the feeling I want someone to have when they see it is [the feeling].

Before you write anything, ask me up to five questions, one at a time, about the things that will actually change the picture: the setting and the moment in time, who or what the subject is and what state it is in, the palette, whether text will sit on top of it and where, and any real reference I have in mind. Do not ask me about resolution, style keywords, or which model I am using.

Then write the brief in four labelled parts. THE SCENE: the situation, the setting, and the moment, written as if you were describing a still from a film. THE LIGHT: the direction, quality, colour temperature and contrast of the light, plus the practical sources visible in frame, and one sentence on why this light suits the feeling I asked for. THE FRAME: the camera angle and height, the distance and the equivalent lens, where the subject sits in the composition, and where the negative space is and what it is for. THE DETAILS: the three or four specific physical details that make the scene believable, such as condensation, wear on a surface, a cable that is slightly out of place, and the things that must NOT appear.

Finally, output the whole brief as one continuous prompt paragraph I can paste into an image model, with no headings and no bullet points, ordered subject first, then setting, then light, then framing, then details. Keep it under 150 words and cut every adjective that does not change what the model would draw.

Lighting words that actually do something

You do not need to memorise these. You need to recognise them when Claude uses them, and to be able to say "no, harder light" instead of "make it better". Six controls cover almost everything.

Framing words that decide where the eye lands

Iterate in the same chat, do not reroll

Here is the habit that separates people who get good images from people who generate forty and pick the least bad one. When something looks off, do not regenerate and hope. Go back to the same Claude chat, describe what is wrong in ordinary language, and let Claude rewrite the brief with the fix built in. Claude still has the whole brief in context, so it changes the one thing rather than rolling the dice again.

1. 👀 Name what is wrong in plain words. "The light is too even", "her hands look wrong", "it reads like a stock photo", "the empty space is on the wrong side". You do not need the technical term. Naming the feeling is enough for Claude to translate it.
2. 🔧 Ask for one change at a time. Change lighting or framing or palette, not all three. If you change everything, you learn nothing about what caused the improvement, and you cannot repeat it next week.
3. 📊 Ask for a controlled set. Have Claude write three versions of the brief that differ in exactly one variable, so you can compare like for like instead of comparing chaos to chaos.
4. 📁 Keep the ones that worked. Paste your best final briefs into one document. That document becomes your house style, and it is the raw material for the skill in the next section.

Here is what came back from the brief you wrote, and here is what is wrong with it: [describe what you see in plain language, for example the light is too flat, the subject is too centred, it looks like a stock photo, the mood came out cheerful when I wanted quiet].

Diagnose it before you rewrite. Tell me which specific instruction in your brief most likely caused that, and what the model probably defaulted to where the brief was silent. I want to understand the cause, not just get a new version.

Then rewrite the brief changing only the thing we are fixing, keeping every other decision identical, and give me three variants that differ in exactly one variable so I can compare properly. Say what that variable is in one line above each, then give me the three paste-ready prompt paragraphs.

Stop pasting it, install it

Once you have used the director prompt a few times you will notice you keep giving the same corrections. That is a signal. Turn it into something Claude loads by itself so you never paste the long prompt again.

1. 📚 Collect your winners. Pull together five or six briefs that produced images you actually used, plus the corrections you keep repeating. This is your taste, written down for the first time.
2. 🗂️ Or start with a Project. If you are not ready for skills, create a Claude Project, paste the director prompt and your house rules into the project instructions, and every chat inside it starts as your creative director. This works on any plan that has Projects.
3. 🛠️ Then write the skill. Use the prompt below to turn the whole thing into a reusable skill with a name you can trigger.
4. ➕ Install it. In the Claude app go to Customize, then Skills, then Add, then Upload a skill, and upload the zipped folder. In Claude Code, drop the folder into your ~/.claude/skills/ directory. Skills are available on every plan from Free through Enterprise.

Turn my creative direction process into a reusable Claude skill so I stop pasting the long director prompt every time.

Here are the briefs that produced images I actually used: [paste five or six]. Here are the corrections I keep repeating: [paste your list, for example "never flat front lighting", "always leave negative space on the left for a headline", "warm palette, no teal and orange"]. Read them and tell me what my consistent preferences are before you write anything, including the ones I have not noticed myself.

Then write the SKILL.md. Give it a short lowercase-hyphenated name, a description that says it writes image generation briefs and triggers whenever I ask for an image prompt, and instructions that make you interview me first, then produce the scene, light, frame and details brief, then the paste-ready paragraph. Bake my preferences in as defaults I can override rather than as hard rules. Finish by telling me how to save it as a zipped folder I can upload.

The brief is model-agnostic

That is the point of working this way: the brief is the asset, and it outlives whichever tool is fashionable this quarter. A few honest notes on where it lands well.

The 4 things that quietly ruin an image