Design#Open source#Agents#Agent Skill#Logo design#Mascot design#Branding
ip-as-logo: make the agent draw mascots that survive 32×32
An MIT-licensed open skill that puts hard limits on how an agent draws IP mascots: 4–7 large shapes per silhouette, three colors at most, legible at 32×32. It pitches three directions first, then generates six candidates in one batch. For developers tired of AI mascots that turn to mush at icon size.
Skill details
npx skills add s1dashu/ip-as-logo-skill --skill ip-as-logoAsk an image model for a product mascot and you tend to get something lovely at full size — rich detail, careful lighting — that dissolves into an unreadable blob the moment it becomes an app icon. ip-as-logo, from the GitHub repo s1dashu/ip-as-logo-skill, takes the opposite approach: it doesn’t add capability to the model, it adds discipline, replacing the “one more detail” instinct with hard complexity limits.
The order it works in
Once installed, the agent reads your project’s README, docs, and design tokens to work out what the product is; if the context is thin it asks you one consolidated round of questions, and only one. It then proposes three directions, one line each: the subject, the product trait it stands for, how the silhouette is built. Generation starts only after you nod — six independent candidates in one batch, two variants per direction, half emerging from the lower-left corner and half from the lower-right so you can compare. Everything comes back as generated: no filtering, no scoring, no automatic retries. If a candidate disappoints, you ask for another draw.
What it constrains
- Complexity: one dominant silhouette built from 4–7 large basic shapes; any shape that can’t explain its role gets merged or deleted.
- Color: exactly three semantic colors per image — two IP base colors plus one solid background — and even the facial marks must come from those two.
- Legibility: the result has to survive 32×32; any feature that vanishes or turns to noise at that size gets enlarged, merged, or removed.
- Composition: the character emerges from the lower-left or lower-right corner and fills 85–95% of the canvas, never the center. Sharp corners, pointed ears, and thin lines are banned; every necessary tip ends blunt and rounded.
- Prompt discipline: the prompt sent to the image model never contains the words logo, app icon, or brand mark — only “a square character image”. The moment a prompt says “logo”, models start piling on detail.
Who it’s for
Developers and small teams who want a mascot without a designer on hand. Two things to know first: the skill ships instructions only, no model — you need an agent wired to a top-tier image generator (GPT Image 2 preferred, with Seedance 5.0 Pro, Nano Banana Pro, or Nano Banana 2 also supported; Codex, Coze, Doubao, Manus, Gemini Apps, and Replit Agent are listed as compatible), and it would rather have you switch models than fall back to SVG. Generation is treated as a one-pass draw with no automatic filtering or retries. If you’d rather skip the install, the project’s site ipaslogo.com keeps a library of ready-made mascots that are free for commercial use. The repo hasn’t been touched since late August, though with a single-document skill there’s not much to rot.