All Ten Of These Ask For A Vector File. Claude Cannot Draw At All.
“STOP ASKING CLAUDE FOR ‘A LOGO’.” The caption crosses out three lazy prompts — design my logo, make me a logo, create a professional logo — and makes a fair point: “Generic prompts = generic results. If you want premium, distinctive, agency-level branding, give Claude a real creative brief.” Then ten briefs, one per comment, and they are properly written. Negative-space symbol, technology mark, fashion wordmark, app icon, mascot, fitness identity, architectural monogram, skincare wordmark, wellness mark, coffee badge.
They are good briefs. They are also addressed to a model that cannot draw.
The source post — “Stop asking Claude for a logo”, BANI Academy. Ten logo briefs, each posted in full in the author’s own comments, every one ending in a list of things to avoid.
What the documentation says
From the FAQ at the bottom of Anthropic’s own vision documentation, in full:
“Can Claude generate or edit images?”
“No, Claude is an image understanding model only. It can interpret and analyze images, but it cannot generate, produce, edit, manipulate, or create images.”
The whole of that page is about sending images to Claude — base64 blocks, URLs, the Files API, size limits, how many you can attach. There is no direction of travel the other way.
So ten image-generator briefs have been pointed at the one model in the stack that does not generate images. Paste them into Claude as written and you will get a thoughtful written description of a logo, or a polite note that it cannot make one.
But the headline is accidentally right
Here is the part that turns this from a gotcha into something useful. Read what the ten briefs keep asking for:
- “Make it instantly recognizable at 32px”
- “Keep it scalable and memorable” — and “prioritize simplicity, memorability, scalability”
- “Highly recognizable on apparel, profiles, and signage”
- “Keep it timeless, editorial, distinctive, and packaging-ready”
- “Optimize for small packaging and labels”
- “Use consistent line weight and perfect optical balance”
- And, in every single one of the ten: “Show symbol-only, horizontal, and monochrome versions”
Every one of those is a property of a vector file. A mark that has to survive at 32 pixels and on a shopfront is an SVG by definition, not a PNG at whatever resolution the generator felt like. “Consistent line weight” is a stroke attribute. A monochrome version is a fill swap. A lockup set is three artboards from the same paths.
And Claude writes text. SVG is text. Ask it for the logo as SVG source and you get an actual vector: open it in a browser, in Illustrator, in Figma, in Inkscape; scale it to a billboard; change one fill to get the monochrome version; edit a single number to adjust a gap.
So the card’s headline is correct for a reason it does not give. Stop asking Claude for a logo. Ask it for SVG.
Diagram: what the briefs ask for, and what each format can do
Take the requests literally and sort them by whether the file type can even honour them.
| What the brief asks for | A generated PNG | An SVG Claude wrote | Why |
|---|---|---|---|
| “Scalable” | No — fixed pixels, blurs when enlarged | Yes — that is the definition | A vector stores shapes, not a grid |
| “Recognizable at 32px” | Only if generated at 32px, where it will be mush | Yes — render at any size | The same paths redraw at every scale |
| “Monochrome version” | A second generation that will not match the first | Yes — change the fill values | One file, two colour settings |
| “Symbol-only, horizontal, full” | Three generations that drift in weight and spacing | Yes — the same paths, rearranged | A lockup set comes from one source |
| “Consistent line weight” | Approximate, and uncheckable | Yes — stroke-width is a number you can read |
You can audit it, not just squint at it |
| “Packaging-ready” / print | No — screen-resolution raster | Yes | Printers want paths |
| Edit it afterwards | Regenerate and hope | Yes — move one point | This is the whole argument |
One caution so this is not oversold. An SVG a language model writes is a first draft, not a finished identity. Curves will be clumsy, optical spacing will be slightly off, and the kerning on a wordmark will need a designer or an hour of your own fiddling. But it is a first draft in the right format, with coordinates you can nudge — which is a categorically better starting point than a 1024-pixel picture of a logo you cannot open.
All ten briefs, and the one thing each is really asking for
| Brief | The signature move | The vector-shaped ask |
|---|---|---|
| Negative-Space Logo | Growth and upward movement “without a literal arrow” | “Scalability” + symbol-only, full, monochrome |
| Technology Symbol | Three interlocking rounded rectangles with meaningful negative space | “Keep it scalable” + symbol-only and horizontal |
| Fashion Wordmark | One seam, clasp or thread detail hidden in a single letter | “Packaging-ready” + wordmark and initial-mark |
| App Icon | The first letter turned into a symbol with one meaningful cutout | “Instantly recognizable at 32px” + icon, wordmark, monochrome |
| Friendly Mascot | Exactly three colours and a silhouette you can read | Mascot-only and full-logo versions |
| Fitness Identity | Two forward-leaning bars with a deliberate gap | “Apparel, profiles, and signage” + symbol, horizontal, monochrome |
| Architectural Monogram | Shared strokes and floor-plan-inspired negative space | “Consistent line weight” + monogram-only and full |
| Skincare Wordmark | A droplet detail integrated into one letter | Full and monochrome versions |
| Wellness Identity | An open circle from two curves, with controlled asymmetry | “Scalable” + symbol-only and horizontal |
| Coffee Badge | A branch with two leaves and three berries, name curved around | “Small packaging and labels” + flat and monochrome |
The best technique on the card is the last line of every brief
Every one of the ten ends with an Avoid list. That alone is unremarkable. What is remarkable is that each list names the specific cliché of that industry rather than general bad taste.
| Brief | What it forbids | What it is pre-empting |
|---|---|---|
| Fitness | “Dumbbells, lightning bolts, flames, gradients, and 3D effects” | Exactly what a model produces for “gym logo” nine times out of ten |
| Technology | “Robots, circuit clichés, gradients, and 3D effects” | The circuit-board swirl that has been on every tech logo since 2014 |
| Skincare | “Leaf icons, scripts, shadows, and decorative elements” | The leaf. Always the leaf |
| Wellness | “Lotus icons, meditation figures, watercolor, and fragile lines” | Four defaults in one line, including the hairline that vanishes at small sizes |
| Fashion | “Crowns, excessive flourishes, and metallic effects” | The gold crown of a thousand boutique logos |
| Coffee | “Clutter, distressing, and mockups” | The fake-vintage weathered texture |
| Mascot | “Complex scenes, tiny details, and heavy shading” | An illustration where a mark was wanted |
| Monogram | “Tangled overlaps and unnecessary details” | The interlock that becomes a knot |
| App icon | “Tiny details, shadows, and mockups” | Detail that dies at 32px — the thing the brief above it asked for |
| Negative-space | “Clichés, gradients, 3D effects, and mockups” | The only generic one of the ten |
Four of the ten also forbid “mockups”, which is the sharpest detail here. Ask an image model for a logo and it will very often hand you a photograph of the logo — embossed on a business card, printed on a tote bag, glowing on a shopfront. That is a picture of a thing, not the thing. Whoever wrote these briefs has been burned by it and said so in advance.
The one thing missing, and it is the one with legal weight
A logo is unlike every other design asset on these cards: it is a trade mark. It is the one piece of visual work that can be already owned by somebody else, and that can be expensive to discover two years after you printed it on your packaging.
Nothing on the card mentions checking. No brief says “then search for existing marks in this class”, and no model — image or text — can tell you whether a mark is in use, because the registers are not in its training data in any reliable form and it has no way to search them for you.
That is not a reason to skip the exercise. It is one extra step, free, before you commit: search the relevant national register and do a reverse-image search on the mark you have settled on. For a channel or a small business the cost of skipping it is small right up until it is not.
Four honest caveats
- We did not run the briefs. Not through Claude, not through an image model. This is a reading of what the briefs ask for against what the documentation says the named tool can do.
- “Claude cannot generate images” has a date on it. That is Anthropic’s published FAQ, read in October 2026. Products change, and if that changes the first half of this article changes with it. The second half — that a logo wants to be a vector — does not.
- The briefs are good for image models too. Paste them into a generator and you will get far better results than “make me a logo”, exactly as the card argues. Our objection is about the format of the output, not the quality of the brief. On that, the card is right and most cards are not.
- Model-written SVG needs a human pass. Expect clumsy curves and off-balance spacing. It is a draft in the right file type, which is a different and better problem than a picture in the wrong one.
The one-line version
Ten well-written briefs, every one of which asks for scalability, a 32-pixel version, a monochrome version or a print-ready mark — and all of those are properties of a vector file, not a picture. The model they are addressed to cannot produce a picture at all. It can produce the vector.
So keep the briefs, keep the Avoid lists — they are the best thing here — and change one word in the request: not “design this logo” but “write this as SVG”. Then open the file and move the points yourself. That is the same distinction we keep landing on from other directions: what is a record and what is a picture of one, and why a graphic that has to be read needs a font rather than a diffusion model.
Try it in two minutes. Take brief number four — the app icon — fill in the brackets, and add one line to the end: “Return it as a single self-contained SVG, 512×512, no gradients, no filters, with a comment above each path saying what it is.” Save the output as logo.svg and open it in your browser. Then resize the window.
One Comment