Heddy IP: The Skill That Keeps an AI Mascot On-Model Across Hundreds of Images

Most "generate me an illustration" skills treat every request as a blank canvas. Heddy IP, shipped in aitutors.me's public plugin repo alongside its flashcards skill, does the opposite: it treats every request as a constraint-satisfaction problem, where the constraint is "this has to still be recognisably the same owl." That's a narrower, harder, and more useful thing to build than a generic image generator, and it's worth understanding why.

The problem it solves

Heddy is aitutors.me's snowy-owl mascot — front-of-house, not a tutor, not a crisis service, just the visual identity that shows up in blog posts, social graphics, and marketing scenes. The moment a brand wants more than a handful of static logo files, it runs into a real problem: ask an AI image model to draw the same character in a new pose, a new scene, or a new medium, and by default you get a similar-looking character, not the character. Eyes shift shape. Proportions drift. A sticker version and an editorial-scene version start looking like two different owls that happen to share a species.

A brand mascot that drifts stops being an asset and starts being a liability — every inconsistent image chips away at the thing that made investing in a character worthwhile in the first place. Heddy IP is built specifically to prevent that, and its design tells you a lot about what "identity governance" actually requires in practice.

How it enforces consistency

The skill's central move is refusing to generate from a text description alone. Instead, it routes every request through a frozen reference set first — a locked identity anchor image plus supporting sheets covering turnarounds, expressions, poses, and wardrobe. Before any generation, edit, or judgement call about whether an image "looks right," the skill loads that reference and treats it as the binding source of truth, not a suggestion.

That single discipline — reference-first, not description-first — is the difference between "an owl mascot, amber eyes, green badge" (which any model will interpret slightly differently each time) and "this exact owl" (which stays anchored because the model is shown the actual character, not just told about it). The skill names the specific features that must never change regardless of pose or scene: asymmetric amber eyes with the left one larger, a small amber diamond beak, a spark-green star badge, and a particular head tuft. Pose, scene, and medium are all free to vary. Those four features are not.

What it's actually for

The skill routes by task rather than trying to be one undifferentiated "make an image" tool: identity questions and on-model repair when something drifted, content illustration for blog and social posts, extending the IP into new poses or wardrobe or scenes, cutout and transparent sticker generation, sizing for specific social platforms, and — planned for later — short video and reels. Generation itself runs on a Gemini API key supplied by whoever's operating the skill; the skill provides the governance layer on top, not the underlying model.

This is a genuinely different category of skill from most of what fills a marketplace like this one. Where a typical skill teaches an agent a general capability — write better commit messages, refactor a function, summarize a PDF — Heddy IP teaches an agent to be disciplined about a single specific asset, treating consistency as a constraint to defend rather than a nice-to-have. That's a useful pattern to notice even if you'll never generate an owl: any team with a recurring brand character, product mascot, or consistent visual motif is solving the same problem, and "reference-first generation with named must-never-change features" is a transferable answer.

Worth knowing before you go looking for the source

One thing worth flagging plainly: Heddy IP's own licence marks it proprietary — internal aitutors.me brand IP, explicitly not for redistribution. That's why this piece describes the mechanism rather than reproducing the skill's prompts or reference assets. It's listed in aiskill.market's catalogue as a pointer to the real source and install path, the same way a marketplace listing points at an app without hosting its source code. If you want to actually run it, the public repo is the place to install it from — not any copy sitting in a database.

The lesson generalises past the licensing footnote, though. A skill can be worth studying for its architecture even when you can't (and shouldn't) copy its content wholesale. The idea of pinning a frozen reference sheet and naming the exact features an image must preserve is portable to any brand-consistency problem, proprietary source or not.

Read more skill spotlights on aiskill.market.