Claude cannot generate images or video directly — but it is one of the best tools you have for producing them. The workflow is simple: Claude writes the prompts, you paste them into a generation tool, and you get results that actually match what you wanted. This approach beats typing into Midjourney or Runway cold because Claude understands context, style language, and how to structure a prompt for a specific model. It also iterates fast. You describe what you need in plain English, Claude translates it into model-ready language, and you test. No guessing at syntax.
This article covers how to use Claude as a creative director for images and video: what prompts to give it, how to refine output, and how to build a repeatable process for content you actually use.
How the Stack Works
You need two things: Claude, and at least one generation tool. For images, common choices are Midjourney, DALL-E 3 (via ChatGPT or the API), Adobe Firefly, or Ideogram. For video, Runway Gen-3, Kling, Pika, or Higgsfield. Claude works with all of them because it is writing text instructions — the model just receives the prompt you paste in.
The reason to run everything through Claude first: generation tools reward very specific, structured prompts. Most people write vague inputs and get vague outputs, then assume the tool is mediocre. Claude closes that gap by knowing the grammar each model responds to — aspect ratios, style tags, lighting descriptors, camera movement language, negative prompts, and so on.
The Prompt
Give Claude this to generate a ready-to-paste image prompt for any tool you are using. Fill in the brackets with your actual details.
You are an expert prompt engineer for AI image generation. I need a detailed, ready-to-paste prompt for [Midjourney / DALL-E 3 / Adobe Firefly — pick one].
What I want to create: [describe the image in plain language — subject, setting, mood, purpose]
Brand context: [colors, tone, style — or write "none" if this is not brand content]
Usage: [where will this appear — Instagram post, website hero, email header, ad creative, etc.]
Output a single prompt block I can paste directly into the tool. Include subject, environment, lighting, style, mood, and any relevant technical parameters (aspect ratio, version flags, quality settings). After the prompt, give me two variations — one with a more editorial/photography feel, one with a more graphic/design feel. Label each one.For video, use this version instead.
You are an expert prompt engineer for AI video generation. I need a detailed, ready-to-paste prompt for [Runway Gen-3 / Kling / Pika / Higgsfield — pick one].
What I want to create: [describe the scene — subject, action, setting, mood]
Brand context: [visual style, color palette, tone — or "none"]
Video specs: [duration if known, aspect ratio, any camera movements you want — pan, zoom, dolly, static, etc.]
Output a single prompt I can paste into the tool. Include scene description, camera direction, lighting, mood, motion style, and pacing. Then give me two variations — one with more cinematic/film grammar, one with a faster, more social-native feel. Label each one.Prompt-Writing Tips That Change the Results
- Name the tool explicitly. Midjourney responds well to artistic style tags and aspect ratio flags (--ar 9:16 --v 6). DALL-E 3 responds better to full descriptive sentences. Runway and Kling need camera grammar ("slow push in," "rack focus from foreground to background"). Telling Claude which tool you are using gets you syntax that actually works.
- Describe the usage, not just the look. "For an Instagram Reel thumbnail" tells Claude to prioritize contrast and legibility at small sizes. "For a website hero" tells it to think about negative space and readability behind text. This context changes what a good prompt looks like.
- Give Claude a reference point if you have one. Paste in a description of an image you like, or name a photographer, director, or visual style. Claude will incorporate that aesthetic vocabulary into the prompt without copying the source.
- Ask for negative prompts too. Most tools accept negative prompts or exclusion language. Add to the end of your request: 'Also give me a negative prompt listing what to exclude.' This prevents common failures — blurry faces, cluttered backgrounds, stock-photo feel, text artifacts.
- Iterate in Claude before you iterate in the tool. Generation runs cost time (and sometimes credits). Before re-running, go back to Claude and say 'The output was too dark and the subject felt too small — revise the prompt.' Claude adjusts the language; you paste the new version. This is faster than trial-and-error inside the tool.
- Request a shot list if you need a series. If you are producing content for a launch, campaign, or client, ask Claude to write five to ten prompt variations on the same theme — different angles, different moments, consistent visual style. You get a cohesive set instead of unrelated one-offs.
- Use Claude to reverse-engineer a look you like. Screenshot an image or video frame, describe what you see to Claude, and ask it to write a prompt that would produce something similar. This is one of the fastest ways to extract a visual style and apply it to your own content.
A Real Workflow Example
Say you are creating a short video ad for a coaching offer. You want a 15-second clip — confident founder energy, urban setting, natural light, nothing that reads as stock footage. Here is how you run it.
Step one: open Claude and paste the video prompt above, filling in your details. Step two: Claude returns a Runway-ready prompt with two variations. Step three: you paste the one that reads closest to your vision into Runway Gen-3. Step four: if the output misses something — lighting feels off, the motion is too slow — you go back to Claude and describe what you saw. Claude revises the prompt. You run it again. Step five: you have a usable clip in two to four rounds, not twenty.
The same loop applies for images. The difference is the feedback cycle is faster because image generation takes seconds, not minutes.
Prompt Variations to Try
- Brand consistency check: 'I have an existing image I want to extend into a series. Here is a description of it: [describe it]. Write five prompt variations that would produce images with the same visual DNA — same lighting quality, same color temperature, same mood — but showing different subjects or moments.'
- Storyboard prompt: 'I need a 30-second brand video. Break it into six shots. For each shot, write a Runway prompt that covers the scene, camera move, lighting, and duration. The overall tone is [describe it]. The brand colors are [list them].'
- Thumbnail optimization: 'Write a Midjourney prompt for a YouTube thumbnail. The video is about [topic]. The thumbnail needs to stop scroll, use high contrast, and work at 1280x720. The channel visual style is [describe it]. No text in the image — I will add that separately.'
- Product visual: 'Write a DALL-E 3 prompt for a product photo of [product description]. Style: clean studio shot, white or very light background, soft directional lighting, professional e-commerce quality. The product should feel premium. Output the main prompt plus a lifestyle variation where the product appears in a real-world setting.'
- Persona / avatar: 'I need an AI-generated persona image for a brand character — not a real person, clearly illustrated or stylized. The character is [describe personality, role, look]. The visual style should feel [approachable / authoritative / playful / editorial]. Write prompts for Midjourney and for DALL-E 3 separately, since they handle this differently.'
One Thing Most People Skip
After you generate something you like, save the prompt. Not just the image or video — the actual prompt text. Keep a running document of prompts that worked, what tool you used, and what you were going for. Over time this becomes a style library you can hand off to a contractor or build on for future campaigns. Claude can also help you organize it: paste your best prompts in, ask it to tag them by style and usage, and you have a searchable asset that gets more valuable the longer you run it.
Claude works best in this stack as a translator and refiner, not a one-shot oracle. Give it context, react to the output you get, and keep the loop tight. Two or three rounds with Claude will consistently outperform twenty rounds of blind trial-and-error in the generation tool itself.