How to create on-brand visuals with Claude: set up a brand design system, generate code-based artifacts, direct an image model via MCP, keep it on brand.
Claude is a language and reasoning model, not an image generator — it has no native text-to-image diffusion model, and that is a deliberate design choice by Anthropic. So "make on-brand visuals with Claude" is not one task, it is three: generate code-based visuals (SVG, HTML/CSS, charts, QR codes, print-ready PDFs) directly in artifacts, build and enforce a brand design system so every output stays consistent, and direct an external image model through MCP when you need a photographic or illustrated raster asset. Done right, Claude is a genuinely strong art director and design-system engine; done wrong, you ask it for a photo and get an apology.
This walks the practical path: tell Claude what your brand looks like in a form it can actually read, lock that into a reusable system so it does not drift after a few turns, then produce and export the asset. It is the make-the-visual companion to the broader workflow in [how to build an AI creative director with Claude](/how-to/build-an-ai-creative-director-with-claude) — read that for the concept-and-brief side; this is the hands-on production of the visual itself.
The steps
Pick the right lane for the asset you need. Decide first what kind of visual it is, because that decides the method. A logo, chart, diagram, social-card layout, landing page, QR code, or print-ready flyer is a code-based visual — Claude writes SVG, HTML/CSS, or a React component and renders it live in an artifact, exactly the kind of thing it is good at. A photograph, a styled illustration, or anything photoreal needs a real diffusion model, which means directing an external image generator (step 5). Naming the lane up front saves you from asking Claude to paint a photo it structurally cannot make.
Give Claude your brand in a form it can read. Claude can read images, PDFs, and design references, so hand it the real material: upload your brand guidelines PDF, your logo, screenshots of existing on-brand posts, and your exact hex colors, font names, and spacing rules as text. If you use Claude Design, point it at a codebase, design file, or a DOCX/PPTX so it extracts your real components, colors, and type. The more concrete and machine-readable the brand input, the less the output drifts toward generic.
Lock the brand into a reusable system, not a one-off prompt. A brand pasted into one message decays after a few turns. Make it persistent instead. In Claude, build a reusable Skill that carries your palette, typography, and layout rules and applies them to every artifact automatically — Anthropic's own public brand-guidelines Skill (its dark/light base colors, orange/blue/green accents, Poppins headings with Lora body, falling back to Arial and Georgia) is the working template to clone and replace with your own values. In Claude Design, an admin can lock a team to a single approved design system so every build stays on brand regardless of who is prompting.
Generate the code-based visual in an artifact. Now ask for the asset in plain language: a square social graphic as SVG, a carousel-slide set in HTML, a brand-colored chart, a business card with a rounded-module QR code, or a print-ready flyer with bleeds, crop marks, CMYK color, and 300 DPI. Claude renders it in the conversation (or on the Claude Design canvas) and, when it is working against a locked design system, checks its own output against that system and corrects before you see it. Iterate in the same thread — "tighten the spacing," "use the accent on the header only."
For raster images, connect an image model via MCP and let Claude direct it. When you need an actual photograph or painted illustration, give Claude an image generator over the Model Context Protocol — GPT Image, DALL·E 3, Flux, Ideogram, or Stable Diffusion — and generate inside the chat. Claude's job here is direction, not rendering: it writes precise, brand-consistent prompts from your style guide, critiques the returned image against your palette and mood, and re-prompts until it lands. The pixels come from the external model; the brand discipline comes from Claude.
Audit every output against the brand guide. Do not trust the first pass. Ask Claude to audit the asset against your uploaded guidelines — "check this against the brand PDF: are the colors, fonts, and logo clear space correct?" — since the same vision ability that reads a reference can grade an output. In Claude Design, use on-canvas editing to comment on any element, edit text directly, or nudge spacing and color with the sliders. This audit step is what separates on-brand from approximately-on-brand.
Export to where the visual is actually going. Finish by getting the asset out in a usable format. Code-based artifacts download as SVG, HTML, or PDF. Claude Design exports to PowerPoint, PDF, and HTML and pushes into tools like Canva, Figma-style editors, and others; Claude Slides and Docs build editable branded decks and documents in-conversation. Expect to do light cleanup on handoff — an exported deck or Canva file rarely lands pixel-perfect — but the structure and brand styling carry over.
Common gotchas
Asking Claude for a photo with no image model connected gets you an SVG approximation or a written description, not a photograph. Connect an MCP image generator first, or accept that the lane is code-based visuals.
A brand PDF dropped into one message is not enforcement. Without a reusable Skill or a locked design system, the output drifts off-brand within a few turns — build the system in step 3, don't re-paste the brand each time.
Code-based visuals are excellent for layout, type, charts, and logos, and wrong for anything photoreal. SVG cannot fake a lifestyle photo; don't force it.
An MCP image model's quality, and its brand-safety, are the external model's — Claude directs the prompt but can't guarantee an exact logo, a specific face, or text rendered correctly inside the image.
Claude Design's availability, model, and export integrations have shifted since launch; check what your account actually has rather than assuming a feature from a recap article exists for you.
It stops at the asset. Claude makes the visual; it does not caption it for each platform, schedule it, or publish it — that is a separate job and a separate tool.
Where Kompozy fits
Everything above is a single-asset loop: you set the brand, describe one visual, audit it, export it, and start over for the next one. That is the right shape when the deliverable is a bespoke landing page, a one-off deck, or a design system — work Claude is genuinely good at. It is the wrong shape when the deliverable is this week's content: thirty on-brand assets across formats and platforms, every week, without thirty trips through the canvas. That throughput gap is exactly where [Kompozy](/) fits, and the fit is production, not direction.
The difference is that Kompozy generates the whole set from one brief instead of one artifact from one prompt. A [Persona Brief](/glossary/persona-brief) holds your voice, palette, and banned words once and governs every output, [HyperFrames](/glossary/hyperframes) renders brand-exact layouts pixel-for-pixel so a carousel or quote card comes out on-brand by construction rather than by you grading each pass, and a Gemini face-lock keeps an AI influencer's face identical across a month of posts — the net-new formats Claude's code-based lane and its MCP image models simply don't produce: [Persona Shorts](/glossary/persona-shorts) and avatar video, [Clipped Shorts](/glossary/clipped-short), Carousel Posts, Photo Posts, Quote Graphics, blogs, newsletters, across [18 formats](/glossary/output-buckets) in one pass. Then [Autopilot](/glossary/autopilot) fans the approved set across the eight social platforms plus blog and email behind a per-post review gate — the make-it-to-posted step this tutorial ends just short of.
Honest split: to design one perfect branded artifact or stand up a locked design system, stay in Claude — it is built for that and Kompozy is not a design canvas. To turn an on-brand look into a standing stream of finished, scheduled, multi-platform content, that is Kompozy's job. Starter ($199/mo, 5,500 credits) fits a solo creator holding one brand steady; Pro ($499/mo, 18,000 credits) suits a brand or agency running a daily multi-format cadence; Enterprise is custom.
Frequently asked questions
Can Claude generate images on its own?
Not as photographs. Claude has no native text-to-image (diffusion) model — that is a deliberate design choice by Anthropic. What it can do is create code-based visuals (SVG, HTML/CSS, charts, diagrams, QR codes, print-ready PDFs) that render in the conversation, and direct an external image model connected over MCP — GPT Image, DALL·E 3, Flux, Ideogram, Stable Diffusion — so you can generate raster images inside the chat with Claude writing the prompts.
How do I keep Claude’s visuals on brand across many outputs?
Make the brand persistent instead of re-pasting it. Build a reusable Skill that carries your colors, fonts, and layout rules and applies them to every artifact, or in Claude Design have an admin lock the team to a single approved design system. Anthropic's own brand-guidelines Skill is a public template: it applies a fixed palette, Poppins/Lora typography, and font fallbacks to every generated artifact — clone the pattern and swap in your values.
What is Claude Design and how does it help with brand visuals?
Claude Design is Anthropic's visual workspace where you describe a visual and Claude builds a first on-brand version. It imports DOCX/PPTX/XLSX, design files, and codebases so it builds with your real components, checks its output against your design system and corrects before you see it, and exports to PowerPoint, PDF, and HTML plus integrations with tools like Canva. Admins can lock a team to one approved design system so everything stays on brand.
Can Claude make a branded carousel or pitch deck?
Yes. A carousel is a code-based layout job Claude handles well as HTML or SVG slides, and Claude Slides builds editable branded decks in-conversation that export to PowerPoint. Both stay on brand if you work against a locked design system or a brand Skill rather than a one-off prompt. Expect light cleanup on export — slide decks rarely land perfectly formatted — but the brand styling and structure carry over.
Can Claude publish the visuals it makes to social media?
No. Claude produces and exports the asset; it does not reframe it per platform, schedule it, or publish it to your social accounts. Getting an on-brand visual from "made" to "posted everywhere on a cadence" is a distribution job that lives outside Claude — that is the gap a publishing engine like Kompozy fills.