Logo Design Skill: Turn Claude and AI Agents Into a Disciplined Logo Designer

Ask an AI agent for a logo and you usually get clip art with your company name under it: a swoosh, a lightbulb, a hexagon that means nothing. Logo Design Skill exists to stop that.

It is a free, MIT-licensed skill that turns Claude, or any agent that supports the open Agent Skills format (Gemini CLI, Codex CLI, Cursor, GitHub Copilot and more), into a disciplined identity designer that works from the first brief to production-ready SVG files and brand guidelines. The repository was created on 26 September 2026 and passed 1,400 stars within its first week on GitHub.

Nothing here runs as an app. The skill ships as a SKILL.md file, a shelf of reference documents and dependency-free Python tools that your agent loads as standing instructions. A note on the GitHub language label, which says HTML: the substance is markdown and Python, and the HTML label comes from the local gallery page that browses the reference library. Clone this expecting a web project and you will find a design school for your agent instead.

Every category has its clichés, and AI agents reach for them first. Food delivery gets a fork, a chef hat or a scooter. Security gets a padlock or a shield. Health gets a cross or a heartbeat line. The README's briefs read like a banned-symbols list on purpose: Kiln, a specialty coffee roaster, is told "not rustic cliché"; Keyfort, a password manager, gets "no padlocks, shields, keyholes, fingerprints or chain links"; Mediora, a telehealth app, gets "no crosses, stethoscopes, hearts or pulse lines".

The skill's first move is research: it searches its 1,400-logo reference library to map what the category already looks like, so the concepts it builds can avoid both the clichés and accidental look-alikes. That step alone separates it from a prompt that says "design a modern logo" and hopes.

The second move is craft that most people never see. Logo work lives in details like optical corrections (overshoot, the bone effect, irradiation), geometric construction and snapping diagonals to exact angles. The skill's audit caught a 60.8° diagonal in an early concept and an off-centre symbol before anything was shown to the client. Our Color Converter is the tool you will reach for when the skill hands you a brand palette and you need it in hex, RGB and HSL without thinking.

The skill runs a fixed pipeline, and each stage has a job.

• Brief: the agent starts from your stated assumptions or runs a discovery question bank: word mapping, choosing the right mark type (wordmark, letterform, pictorial, abstract, emblem, mascot and the rest), and a decision guide for which fits.
• Research: it queries the reference library for the category, studying construction and conventions, and checking that its ideas do not echo an existing mark.
• Concepts: 8 to 12 one-line ideas, of which 3 get built in SVG.
• Test and refine: every concept goes through svg_audit.py and a test sheet before you ever see it: 96 and 32 px side by side, one-colour and reversed versions, then a size ladder down to 16 px, plus backgrounds, treatments, contexts and a competitor shelf test.

The reference documents cover the full discipline on demand: the twelve principles, mark types, visual techniques (geometry, grids, balance, optical corrections, negative space), colour, typography, the design process, SVG construction, a testing checklist, presentation and delivery, identity systems, redesigns and critique. The agent loads what the task needs and nothing else, which keeps the context lean.

The skill always stops at the checkpoint. It shows the concepts as one overview image with a recommendation and offers the full kit. Nothing else is produced until you choose a direction, because the kit (colour, lockups, presentation board, icons, guidelines) is most of the work and only makes sense for an approved idea.

This is the design decision that makes the whole thing work. An agent left to run end to end will polish the wrong idea beautifully. Forcing a human decision at the concept stage is what keeps the labour pointed at the right target.

The 28 example brands prove the checkpoint earns its place. In four of them the client picked a different concept than the one the skill recommended, and the skill refined that choice before colouring it: Tinkertrail's client took the letterform over the recommended snail because it reads at 16 px; Hearsay's client took the pull-up chair over the recommended ear. That is exactly what the checkpoint is for. A recommendation you cannot overrule is not a recommendation.

The library is the skill's unfair advantage. More than 1,400 SVG files covering roughly 1,200 brands, each visually classified by mark type, technique, geometry, subject, typography, mood and industry. The breakdown reads like a design textbook's table of contents: abstract 24%, combination 21%, pictorial 20%, letterform 15%, wordmark 10%, emblem 4%, mascot 4%, lettermark 3%. Median 2 colours, three quarters of the library uses 3 or fewer, gradients appear in 19%, and 141 files are flagged as exemplary teaching examples.

It is searchable from the command line (python3 scripts/search_library.py --technique negative-space --exemplary, or --industry payments-fintech --summary) and browsable in a local gallery page. The stated purpose is study: learn construction, map category conventions, avoid look-alikes. Never to copy. The library guide says it plainly, and the TRADEMARKS.md file draws the legal line (more on that in the limitations section).

For comparison of scale in this ecosystem, Anti-Slop ships a rulebook where this ships a school. Both are agent skills; one filters output, the other teaches craft.

Everything is plain Python with no dependencies beyond the standard library (Python 3.8+), so the tools run anywhere your agent runs.

• svg_audit.py turns craft rules into checks: viewBox and aspect, colour count, anchor points against library medians, near-miss angles, tiny details under 1/48 of the canvas, centring. It ends with a production-readiness score out of 100. The README's own example scores a recommended lockup at 99 and a rejected emblem at 76, with the warnings spelled out.
• preview_sheet.py builds an HTML test sheet: sizes, backgrounds, treatments, contexts and the competitor shelf test.
• concept_sheet.py renders the one-image concept overview the checkpoint shows you.
• presentation_board.py builds client presentation boards with six industry-specific mockups per concept.
• render_png.py and export_variants.py take the final SVG to transparent PNGs at exact sizes, a favicon.ico, black/white/mono/square variants and the complete favicon, app-icon and web-manifest set.

PNG and ICO export uses whatever renderer is available: cairosvg, rsvg-convert, Inkscape, a Chromium-based browser or macOS Quick Look. If you hand-edit the exported SVG afterwards, running it through an HTML minifier keeps the file lean, and embedding it as a data URI is a one-liner with Base64 encode.

Twenty-eight fictional briefs, from quiet luxury and neon festivals to B2B SaaS, fintech and health, each run end to end with the skill. For every brand you see exactly what the skill shows at the checkpoint (greyscale concepts at true 64, 32 and 16 px sizes plus a recommendation), followed by a colour preview of the chosen direction on mockups picked for that industry.

A few of the strongest: Kiln, a specialty coffee roaster, where the recommended mark is a K whose leg is the kiln's arched firing mouth, with the arch returning as the n of the wordmark. Zestly, a food delivery app, where a citrus wedge flips into a cheeky grinning face, and the skill flags its own honest risk: some people read the wedge as watermelon. Maison Orvelle, a luxury fashion atelier, where a Didone M holds an O like a pendant at a V-neckline, and the craft pass killed an M-in-a-circle for sitting too close to a famous transit sign. Keyfort, a password manager, where one unbroken wall winds inward into a Greek key that is also a fortress with only one way in.

The later batches deliberately push vivid, saturated palettes while still passing the one-colour and 3:1 contrast checks, which is the hard combination. Colour that only survives at full saturation is decoration; colour that survives one-colour is identity.

Claude Code users get the recommended route, a plugin marketplace install:

/plugin marketplace add kaankiziltug/logo-design-skill
/plugin install logo-design@logo-design-skill

Manual install is a copy: git clone https://github.com/kaankiziltug/logo-design-skill.git then cp -r logo-design-skill/skills/logo-design ~/.claude/skills/logo-design for personal use (or .claude/skills/logo-design per project). Claude.ai and Claude Desktop users download logo-design.zip from the Releases page and upload it under Settings, Capabilities, Skills; if the upload has a size limit, logo-design-lite.zip carries everything except the SVG files themselves.

Gemini CLI, Codex CLI and other agents use the open Agent Skills format, so it is the same drill: clone once, copy the folder into your agent's skills directory (~/.gemini/skills/logo-design, ~/.codex/skills/logo-design, and so on), start a new session and ask.

Then just ask, in plain language:

*"Design a logo for Harbor, a savings app for first-time savers. It should feel calm and safe." "Here's our current logo (logo.svg), critique it and suggest a refresh." "Give me three wordmark directions for a specialty coffee roaster called Kiln." "Turn this symbol into favicon, app icon and one-colour versions, plus a one-page usage guide."*

Two operational notes. The skill works best with a model that can view images, because it renders its own drafts to PNG and checks them before showing you anything. And like all skills, it loads at session start, so open a new session after installing.

Two caveats, both stated in the repository itself.

First, the license split. The skill text, scripts, templates and catalog data are MIT. The 1,400 SVG logo files in assets/library/svg/ are trademarks of their respective owners, included for reference and education only, and are not covered by the MIT license. Study them, learn from them, do not ship them. The skill is built to prevent copying, but the legal responsibility for what you publish stays yours.

Second, vision matters. The skill renders its own drafts to PNG and checks them before showing you anything, which is how it catches things like a letterform that reads as a different letter at 16 px. That loop only closes with a model that can view images. Run it on a text-only model and you still get the SVGs and the written reasoning, but you lose the self-review pass, and the self-review is where half the craft lives.

The broader limitation is the honest one for any design tool: it executes a process with unusual discipline, but the brief is still yours to write well. A vague brief gets you disciplined clichés.

The repository was created on 26 September 2026 and passed 1,400 stars with 79 forks in under a week; the latest push landed on 30 September, two days before this page went up. It carries a badge for #14 on The New 100 agent leaderboard, and the test suite runs on every push via GitHub Actions.

The growth curve is the usual early hockey stick, which is either a community forming or a trend passing, and a week is too early to say which. The durable signals: the documentation is enormous and specific (the README alone runs close to 500 lines), the tools are dependency-free and tested, the 28 worked examples show the process rather than describing it, and the project is unusually willing to state what it will not do.

Good fit if you need a real logo for a side project, startup or open-source tool and want actual craft (geometry, optical corrections, testing at 16 px) instead of clip art with your name under it.

Poor fit if what you need is a human designer's taste, negotiation and accountability, or you were hoping to lift marks out of the reference library. It is built to stop exactly that.