The mandatory checkpoint is what makes this AI logo skill actual design methodology

What this repo actually encodes is a design methodology, not just a code tool. The skill works by giving Claude — or any agent that speaks the open Agent Skills format — a structured workflow: discovery brief, word mapping, mark-type selection, geometric construction, optical correction, colour, lockups, testing, and delivery. The agent doesn't freewheel; it stops at a mandatory checkpoint, shows three concepts as a single overview image with a recommendation, and waits. Nothing else ships until you pick a direction. That constraint is doing real work: it prevents the agent from generating a full brand kit for a concept nobody approved.

The reference library of over 1,400 real-world SVG logos is the most interesting infrastructure here. Each file is classified by mark type, technique, geometry, subject, typography, mood, and industry. The library exists to study construction conventions and flag look-alikes — explicitly not to copy. The breakdown is telling: abstract marks lead at roughly a quarter of the library, combination marks and pictorial marks follow closely, and mascots and lettermarks together account for under ten percent. That distribution reflects actual professional practice more honestly than most design tutorials do.

The Python tooling is dependency-free against the standard library, which matters for agent environments where installing packages is friction. svg_audit.py turns craft rules into automated checks — near-miss angles, anchor-point bloat, tiny details below a size threshold, centring — and produces a production-readiness score. The audit example in the README is instructive: a recommended lockup scores 99/100 with 205 anchor points; a rejected emblem scores 76/100 with 374 anchor points, flagged as exceeding the 95th percentile of comparable reference logos. That's a concrete, reproducible signal, not a vibe.

The eighteen example runs are genuinely useful documentation. Each brief includes explicit negative constraints — no forks or chef hats for the food delivery app, no padlocks or shields for the password manager — and the skill's reasoning about rejected concepts is preserved. The Maison Orvelle run dropped a shared-foot "LL" because it read as ORVEILE; the Keyfort run dropped a star-fort concept because it read as a shuriken. These aren't edge cases; they're the actual work of logo design, and encoding that reasoning as part of the skill's process is what separates this from a prompt that says "make me a logo."

The format compatibility is broad: Claude Code via marketplace or manual install, Claude.ai via zip upload, Gemini CLI, Codex CLI, Cursor, GitHub Copilot. The skill is plain Markdown instructions plus plain Python, so any agent that can read a SKILL.md and execute scripts can use it. The one real constraint the README flags honestly: the skill works best with a model that can view images, because it renders drafts to PNG and checks them before showing anything.