A skill that makes Claude write image prompts the way an art director briefs a photographer, in the dialect of the tool you use: Midjourney, GPT Image in ChatGPT, Nano Banana in Gemini, Adobe Firefly, Ideogram or FLUX. It starts with the image's job and where it will run, so the shape and pixel size are right before a word is written. It builds a shot card of seven decisions, such as camera height, light direction and where the headline space sits, then turns it into a prompt with the right settings and references. After every result it reads the image, ranks what is wrong and writes one fix.
Inside are four reference files. A tool sheet covers all six tools as checked on 17 October 2026: current models, parameters and their ranges, aspect ratios, sizes, reference limits and editing routes, each with its source. A visual vocabulary gives the camera, light, composition and colour words that steer an image, with numbers, and the shape for each destination. A fix library sets the order for reading a result and lists faults with their fixes. A worked example follows a fictional brand from brief to locked prompt. The skill makes one change per fix, never invents label text, and keeps artist names and real likenesses out of every prompt.
Claude cannot make the images. You run each prompt in your own tool and send back the file. Generators still redraw labels and misspell small text, so the skill tells you when to place the real product photo or set the words as live type, instead of fighting the tool for an afternoon. Tool features change every few weeks, and your screen wins when it disagrees with the sheet. Turn on code execution so Claude can measure results, work out crops and draw a layout guide to attach. Turn on web search so it can check a tool's help page. Send downloaded files, not screenshots, because previews hide spelling mistakes.
Claude skill · version 1.0
AI image prompt director
Checking your access.
In the download · 6 files
- SKILL.mdThe method Claude follows, in six stages: brief, direction, prompt, variations, read and fix, lock and hand over.
- README.mdHow to install it in claude.ai or Claude Code, how to start, and the licence.
- references/worked-example.mdOne hero image for a fictional outdoor brand, worked through all six stages in Midjourney and ChatGPT, with three fixes and one rolled back.
- references/tool-sheets.mdMidjourney, GPT Image, Nano Banana, Firefly, Ideogram and FLUX as checked on 2026-10-17: models, parameters and ranges, sizes, references, editing routes, rights notes and sources.
- references/visual-vocabulary.mdShapes and pixel sizes per destination, and the camera, light, composition, palette and style words that steer an image, with numbers.
- references/fix-library.mdHow to read a result in a fixed order, when to edit or regenerate, fault tables with fixes, and when to stop and change route.
Install it and begin
01Install it in Claude CodeTerminal mkdir -p ~/.claude/skillsunzip ~/Downloads/ai-image-prompt-director.zip -d ~/.claude/skills/# Start it with /ai-image-prompt-director. If this is your first skill, type /reload-skills once.On claude.ai there is nothing to unzip: upload the zip as it is. Every step is on Install and use a Claude skill.
02Your first message to ClaudeText Use the ai image prompt director skill. I need an image of [WHAT IS IN IT] for [WHERE IT WILL RUN]. I use [YOUR TOOL AND PLAN]. It must show [WHAT MUST BE EXACT, OR NOTHING]. Start with the brief and tell me if my tool fits the job.Attach photos of your product if it appears, and your brand colours or guidelines. Turn on code execution so Claude can measure each result and work out crops for other sizes.
Worked example
One image, from brief to locked prompt
- 01Name the image's job and where it runs, then set the shape from that: 4:5 at 1440 x 1800 for a feed ad, 9:16 for Stories with the top 14% and bottom 35% kept clear of words.The shape decides the composition. A tall frame cut out of a wide one almost always loses something that mattered.
- 02Check the tool against what must be true before writing a word.If the label has to be readable, a tool that redraws letters will waste an afternoon. Attach the real product photos, or plan to place the real packshot afterwards.
- 03Write a shot card of seven decisions with numbers: subject, camera, light, composition, style, palette and shape.Anything left out is decided by the model, and its default is the average stock photo. "Product fills 60% of the frame height, light from the upper left" survives; "premium" does not.
- 04Write the prompt in the tool's own dialect, with settings and attachments kept outside the prompt.Midjourney reads parameters at the end, Ideogram reads them as ordinary words and FLUX has no negative prompt. The same sentence behaves differently in each.
- 05Run three or four variations that each change one decision: camera, light, layout or finish.Changing one thing at a time shows what actually moves the picture, and gives a client real options instead of near copies.
- 06Read every result in order: truth, brief, physics, craft, then taste. Spell out any text letter by letter.A beautiful image with a misspelt label is a failed image. A fixed order stops taste from hiding a fault.
- 07Fix one thing per pass, and change route after three failed passes on the same fault.Stronger words rarely fix a stubborn fault. A composite of the real product, or live type set in design software, usually does.
What to do it with
- ClaudeFree tier, paid above it
A language model that writes, edits, reads long documents and runs tools, including agents that work inside a codebase.
Bad at It is agreeable by default, so it will improve a bad idea instead of telling you it is bad. Ask it to argue against you and the quality of the answer changes.
- Claude CodeSubscription
An agent that reads a whole repository, edits files, runs the tests and explains what it changed.
Bad at It is fast enough to make a large mistake quickly. Work on a branch, read the diff, and never let it push. Watch the starting mode: on Pro, Max and Team plans it opens in auto mode, which Anthropic's own list says may read a .env file and send the key in it to the API that key belongs to. Switch it off.
- MidjourneySubscription
Generates images from a written description, with the strongest default aesthetic of anything in the category.
Bad at That default aesthetic is the problem as often as it is the draw. It has a house look, it pulls everything towards it, and getting a plain honest photograph of a real product out of it is harder than it should be. Text inside an image is unreliable.
- GPT ImagePay per use
Generates and edits images with unusually good instruction-following, and can hold text inside a picture better than most.
Bad at Faces drift between generations unless you feed a reference every time. Skin can come out plasticky at default settings, and a heavy fidelity setting costs many times more and takes many times longer.
- ChatGPTFree tier, paid above it
General assistant with image generation, data analysis and web access in one window.
Bad at Its default voice has become recognisable enough that readers now spot it, which is a brand problem if you ship it unedited.
- GeminiFree tier, paid above it
Google's assistant, with a very long context window and tight links to their other products.
Bad at Behaviour changes between the app, the API and the versions inside other Google products.
- Google AI StudioFree tier, paid above it
Where Gemini API keys, projects, billing and spend caps are made, with a playground for the models.
Bad at Free-tier prompts are used to improve Google's products, and the billing screens changed several times in 2026.
- Adobe FireflyFree tier, paid above it
Adobe's generative models, trained on licensed and public-domain material and wired into Photoshop and Illustrator.
Bad at Raw image quality trails the leaders. What you are buying is the indemnity and the fact that it sits inside files you already have.
- IdeogramFree tier, paid above it
An image model built around getting words right inside the picture.
Bad at Strong at type, weaker at photographic realism. A poster comes out well, a product on a plinth less so.
- FLUXOpen weights or open source
A family of image models, several with open weights, known for prompt adherence and clean type.
Bad at The open versions need a machine and a person who is comfortable with one. Licence terms differ between the variants, so check which one you are running before it goes anywhere near a client.
Why it matters
Ask a chat for an image prompt and you get a paragraph of adjectives ending in "8k, masterpiece", written for no tool in particular. It ignores the shape the image needs, the settings that matter and what the tool can actually hold, and after a bad result it writes more adjectives. This skill makes Claude decide the picture first, write it in your tool's own terms, and fix results one change at a time, starting with whatever must be true.
- Type
- technique
- Period
- ongoing