Extract Your Visual Style, Then Lock It Into an AI Image Workflow

Imagine a digital artist who uploads six pieces from a botanical-glass series into Claude Design and asks what visual system the work follows. A vague mood summary like "soft, floral, dreamy" would be easy to expect, the kind of thing that sounds nice and changes nothing.

Instead, the answer is a working specification. It includes exact hex values sampled from the artist's own renders, plus the proportions they're used in: 46% rose ground, 25% petal pink, 10% sage and 3% berry. It also reports that the focal point lands 42% from the top of the frame in all six pieces, a habit the artist hadn't noticed. One more line stands out, saying the olive ground behind pieces 04, 05 and 06 is "a second brand hiding in your set."

So there were two brands, swapped back and forth for months and passed off as range. That's the moment extraction becomes useful. It isn't when the tool flatters you but when it names the thing you've been avoiding. What follows is the workflow for the stage after extraction, with that spec as the worked example.

Run the extraction first

Start with the extraction if you haven't yet. Go to claude.ai/design, upload your strongest work, and have it pull out the system before requesting anything new. Pro, Max, Team and Enterprise plans include Claude Design (Free doesn't), and since it's labeled a research preview, some roughness should be expected.

Six pieces works well. With fewer than four, it can't tell a pattern from an accident. With more than ten, the contradictions average out into mush, which defeats the purpose. Choose your best work, not your latest, because you're teaching it what you're aiming for. Start with a short framing message.

Here are 6 pieces from my [series name].

Then send the actual request.

Before you generate anything, extract the design system that's already
in these. I want a working specification, not a description.COLOR: exact hex values sampled from the actual pixels. Group them by
role — ground, mid, accent. Tell me the approximate ratio each one
occupies across the set. Name the deepest value in the whole set.COMPOSITION: aspect ratio, where the focal point sits as a percentage
from the top, subject width as a percentage of frame, margin rules,
where secondary elements are allowed to live.TYPE: what typographic decisions exist, and flag where I've made none.Then — most important — tell me the places this set argues with itself.
Where am I inconsistent? Which inconsistency is doing the most damage?Don't be diplomatic about it.

The last paragraph of that prompt is the one that pays off. Everything above it is documentation, while the final part is diagnosis.

Deal with the contested element

Every extraction flags something as inconsistent. In this case it was labeled outright: a swatch tagged CONTESTED, with a note that the olive backgrounds formed a second brand inside the set.

Three of the six pieces sat on warm rose and three on olive-sage. Both looked good, and that was the trouble. Nothing was broken, so there had never been a reason to fix anything, yet the grid read like two accounts sharing one login.

Most people want to choose a winner and delete the loser, and that tends to be a mistake that costs them good work. The smarter approach is demotion, not deletion. Rose became the primary at roughly 80% of output, the ground that makes the work unmistakably the artist's. Olive turned into a named seasonal mode at 20%, saved for deliberate runs such as a launch, a series or a season, and never alternating frame by frame in a grid.

The spec also produced a conditional rule that would be hard to work out alone: when olive is the ground, the glass interiors must go pearl-and-pink rather than green. Otherwise the whole frame becomes one green mass and the subject stops separating from the background.

Here's a prompt worth borrowing. Open with the flagged item.

You flagged [the contested element] as an inconsistency in my set.

Then ask for a demotion plan instead of a verdict.

Don't tell me to pick one. Tell me how to demote one.Which should be primary, and at what percentage of my output? What
should the secondary one be reserved for — what occasion, what
cadence, what rule?And what has to change inside the frame when I switch to the secondary
mode, so it still reads as the same brand?

Turn the spec into a lock block

Leaving a spec in a document achieves nothing until it becomes something you paste. Converting the extracted rules into a block that opens every image prompt may be the best twenty minutes in the whole process. The compressed version starts with a title line.

STYLE LOCK — botanical glass

and continues with the rules themselves.

Ground: dusty rose field (#E3C0B7 body, #F4E2DA upper wash,
#CFA096 corner vignette). Warm throughout — no cool grey, no true
black, no true white anywhere in frame. Deepest value #4A3A38.Color ratio: ~46% rose ground, 14% pearl, 25% petal pink, 10% sage,
3% berry, 2% warm gold.Green appears ONLY inside the glass — never as a background, never
in the outer field.Berry (#C9576F) is the only saturated color permitted. Flower centres
only. Under 3% of the frame.Backlight bloom is always behind the subject, never in front.Format: 4:5 portrait, 1080 × 1350. One subject, horizontally centred,
55–70% of frame width. Never two subjects.Focal point — the eye, the glass rim, the widest point — sits 38–45%
from the top.Bottom 15% is a wet reflective plane. Non-negotiable; without it the
subject floats and the piece reads unfinished.Petals and butterflies live in the outer 20% margin only. They may
break the frame edge. They never enter the centre.

That block is worth a close look, because it contains no mood words at all. Every line is a numeric constraint, and roughly half of them are bans: green appears only inside the glass, there are never two subjects, and nothing sits in front. Bans carry more weight than wishes. A phrase like "soft and dreamy" fits thousands of images, nearly all of them bland. A ban on true black has one outcome, and that outcome is recognizably yours.

Fix things in the right order

The spec ended with a remediation list in priority order, and the logic behind that ordering is worth stealing even if your rules look nothing like these:

1. Add the handle to every piece, identical in placement, size and opacity. Only one of the six carried it. This was the cheapest item on the list, yet it did the most harm, because work with no attribution never builds on itself and every reshare led nowhere.
2. Declare rose primary and stop making olive backgrounds ad hoc.
3. Lock the light to one warm backlight, a soft cast shadow, and no hard window edges.
4. Standardize the frame at 4:5 and 1080 x 1350, with a wet floor plane in every piece.

The order goes from cheapest and most visible to most expensive. Fixes that are fun, and therefore tempting for taste-driven people, don't come first. The handle took under an hour and changed how every later post performs, while the lighting fix is more rewarding artistically and counts for less.

If your list has no order, ask for one. First, ask for a ranking.

Rank these fixes by damage-per-hour-of-work: how much each one improves
brand recognition divided by how long it takes me to implement.

Then ask which to do first.

Cheapest, highest-visibility first. Tell me which one to do today.

Protect the composition

Layout is the step people skip, since color feels like the creative choice and framing feels like paperwork. Yet in this spec the framing came out as the system's strongest asset, the thing most deserving of protection. The focal point sat between 38% and 45% from the top in all six pieces, and five used 4:5. That sameness is what made the pieces read as one collection, holding the identity together while the colors disagreed with each other.

The frame is therefore the cheapest thing to standardize and the costliest to lose. Fix the aspect ratio, the focal height and the margin rule, and save the color arguments for afterward. There's a side benefit too. With the frame locked, each overlay (handle, caption, price, call to action) gets one permanent home. Here that means a 4% inset, handle top-right, caption bottom-left, and nothing in the center. It stops being a design decision and becomes a template, and that template is why a post takes ten minutes instead of sixty.

Package the system as a Skill

Today the spec sits in a document, and you must remember to paste it. That works until a busy week arrives, after which you're improvising again. Skills fix that. A Skill wraps instructions, rules and reference files so Claude loads them on its own whenever they apply. You define the system once rather than restating it at the start of every session.

Load it with all of it: each hex value and its role, the ratio, the composition and negative rules, the verdict on the contested element and the conditional rule attached to it. Add something nearly everybody leaves out, which is a few rejected pieces, each with a one-line reason. Rejections mark the boundary better than approvals do. "Drifted too olive." "Green leaked into the outer field." "Two subjects, doesn't work." Three of those teach more than thirty approved images. The request to Claude looks like this:

Build me a custom Skill from this design system spec.

Include: the palette with hex values and the role of each, the working
ratio, composition rules, type rules, and every negative rule as an
explicit prohibition.Add a section called "rejected" — I'll give you examples of work that
missed, and you write down the rule each one broke.Then tell me how to install it.

Audit every new piece

Now the system earns its keep. A new piece gets checked against the rules and not against your mood that day. Submit it for review.

Here's a new piece for the series. Audit it against my design system.

Then ask for a strict, rule-by-rule verdict.

Go rule by rule: comply or fail. For each failure, name the rule, what
the piece does instead, and the smallest fix.Then tell me the one thing that's most off-brand about it.Don't grade generously. If it passes everything, say so plainly — but
check the negative rules first.

Running it before anything is posted or packaged costs about ninety seconds, and it exposes the slow drift that stays invisible until the grid looks unfamiliar three months later.

Know which rules can bend

A system is only worth having if breaking it carries meaning. Follow every rule on every piece forever and you end up with a template, which is a different thing from a brand. The test is whether you can say which rule you broke and for what reason. A launch piece made maximalist because it has to stop people mid-scroll is a decision. A piece that simply "came out different" is drift.

List which rules are structural and which are stylistic. Here the frame, the wet floor plane and the handle are structural and never move. Palette ratio and particle density are stylistic and may flex for a launch or seasonal run. Settle that before the temptation to bend something shows up.

What you get beyond consistency

Two benefits show up that you might not expect. The first is speed. Each image prompt starts from a locked block instead of a blank line, so the variable becomes the idea, and no longer whether the piece will look like yours. That's the whole promise of a design system, and it applies to AI image generation just as cleanly as to product design.

The second benefit is that the spec turns into a product. A written-down visual system (palette with roles, composition rules, the prompt block that generates it) is what buyers of prompt packs are paying for. Extraction does more than organize existing work. It digs up an asset that was there all along, undocumented.

Many creators are a single extraction away from discovering they already have a system. It's just unwritten, which means they can't repeat it, teach it or sell it. So upload six pieces, ask what you've been doing, and then take on the hard part: read the contradiction it surfaces and make an actual decision.