A skill that makes Claude hold one look across a whole set of AI images, so picture 40 still matches picture 1. It maps the set first: how many pictures, which kinds, which shapes and which tool makes which. Then it writes the look as a one-page style sheet with numbers and tolerances for light, palette, lens, grain and composition, picks three to five anchor images, and fills in a settings card for your tool with the model version pinned and prompt rewriting switched off. It writes one frozen block of prompt text with slots for only what changes, tunes it on a 12-picture test run, and then checks every batch you upload against the anchors.
Inside are four reference files. The style sheet guide puts light on a clock, gives lens and depth tables, and translates words like premium, warm or moody into figures. The tool file covers style references, seeds and version pinning for Midjourney, GPT Image, Nano Banana, FLUX.2, Ideogram, Recraft, Firefly, Leonardo and ComfyUI, each sourced and checked on 17 October 2026. The checking file has a coded checklist, a Python script that measures warmth, black point, grain and palette against your anchors, a drift log and the image 40 test. A worked example follows a fictional ceramics studio to picture 40. Claude changes one thing per rerun and judges each picture against the anchors, never against the one before.
It cannot make images. It writes the rules, the settings and the prompts for your own tools, and checks the pictures when you upload them. Switch on code execution so it can measure colour and grain, build contact sheets and keep the prompt log as a spreadsheet. Without it, checks are made by eye and colours are given in words. Tools change their defaults and features often, so glance at your tool's settings before a large batch. You get the most from it by bringing pictures you own that already have the look, your real product photos from two angles, and a clear count of what the set needs.
Claude skill · version 1.0
AI style lock
Checking your access.
In the download · 6 files
- SKILL.mdThe method Claude follows: set map, style sheet, reference pack, tool lock, skeleton and calibration, then batches and checks, one stage at a time.
- README.mdHow to install it in claude.ai or Claude Code, how to start, and the licence.
- references/style-sheet.mdHow to write a look as rules with numbers and tolerances: the template, light on a clock, contrast ratios, colour of light, lens and depth tables, composition, the do-not list, mood words translated, illustration sets and versioning.
- references/worked-example.mdA complete lock for a fictional ceramics studio, from the set map and style sheet to calibration runs, batch 4, picture 40 and the drift log.
- references/tool-features.mdStyle references, seeds, version pinning and prompt rewriting for Midjourney, GPT Image, Nano Banana, FLUX.2, Ideogram, Recraft, Firefly, Leonardo and ComfyUI, each sourced and checked on 17 October 2026.
- references/qc-checklist.mdThe coded checklist, the set review, a Python script that measures warmth, black point, grain and palette against the anchors, working tolerances, the drift log, the image 40 test, the finishing pass and when to stop.
Install it and begin
01Install it in Claude CodeTerminal mkdir -p ~/.claude/skillsunzip ~/Downloads/ai-style-lock.zip -d ~/.claude/skills/# Start it with /ai-style-lock. 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 style lock skill. I need [NUMBER] pictures for [PROJECT AND WHERE THEY RUN], made in [YOUR IMAGE TOOL AND PLAN]. [HOW MANY] of them show our real product. I have attached pictures with the look I want and [OUR BRAND GUIDE, OR NOTHING]. Start with the set map and tell me what is missing.Attach three to eight pictures that already have the look, your own or ones you have the rights to, plus product photos from two angles if the product appears. Switch on code execution so Claude can measure colour and grain.
Checklist
Lock check before the first batch
- Every rule on the style sheet has a decision, a number and a tolerance, and the sheet fits on one page."Warm natural light" cannot be made the same way twice, or checked from the picture.
- Three to five anchors share the light, palette and finish but show different subjects, with no product, face or words in them.Anchors that share a subject teach the tool the objects instead of the look.
- Product photos go in as object references only, from at least two angles, and never as style.The product's colour and label are truth, checked on their own, and no grade overrides them.
- The model version is pinned on the tool card and written into every prompt that accepts it.A default model that changes mid-set is a common reason a batch suddenly looks different, and the easiest to prevent.
- Prompt rewriting is off: Magic Prompt, prompt upsampling, prompt enhance and preset styles.Each one adds words you did not choose, and different words every time.
- The variable slots carry only subject, action and set, with no light, time of day, colour or mood words.One word like "morning" or "airy" in a slot is enough to pull a picture away from the set.
- Calibration has passed: at least 11 of 12 test pictures pass every must rule, with one change per run.Production on an untested lock repeats its faults across forty pictures.
- Every new picture is judged against the anchors and picture 1, never against the picture before it.Small shifts add up, and by picture 40 the set has walked away without any single picture looking wrong.
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.
- 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.
- 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.
- 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.
- RecraftFree tier, paid above it
Generates vector and raster artwork, and can hold a defined style across a set.
Bad at Vector output still wants cleaning by hand before it goes into an identity. Style consistency is good, not locked.
- 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.
- LeonardoFree tier, paid above it
A generation product with trained style models, a canvas and finishing tools around the generator.
Bad at Model quality depends heavily on which of its many models you pick, and the naming gives a newcomer no way to guess.
- ComfyUIOpen weights or open source
A node graph for image and video pipelines. Every step of the process becomes a box you can see and rewire.
Bad at The learning curve is real and the community workflows you download are usually undocumented. Debugging somebody else's graph is a day.
- LightroomSubscription
Raw processing, colour and a catalogue, with masks that find skies and people by themselves.
Bad at Its denoise and its presets pull everything towards a current look, and that look dates.
Why it matters
Ask an image tool for forty pictures in one style and each comes back a little different. The light changes side, the whites cool, a prop repeats, and the default model changes halfway through. Nobody notices until the grid is laid out. This skill makes Claude write the look as rules a checker can test, pin every setting that drifts, keep mood and time-of-day words out of the prompt slots, and compare each new picture with the anchors instead of the last one.
- Type
- technique
- Period
- ongoing