Every blog needs a mascot, and mine is a raccoon — specifically, a cute anthropomorphic raccoon wearing rainbow round glasses, a green t-shirt with bold white text, and mismatched Crocs (blue left, yellow right). Here’s the story of how that character evolved from early AI experiments to a cohesive visual brand.
- Origin Story: December 2022
- The Original Pack (v1 — DALL-E era, 2022)
- The Refresh (v2 — GPT Image, 2025)
- The Brand System (v3 — 2026)
- The March 2026 Audit
- The 7 Habits Marathon (May 2026)
- Character Design Reference
Origin Story: December 2022

During a December 2022 time-off, I started experimenting with AI image generation for the blog. The results were charming enough to commit to: a raccoon mascot for every major section of my eulogy.
The first generation used DALL-E (the best available at the time) and produced a set of .png images with inconsistent styles — some 3D-rendered, some semi-photorealistic, some with odd color casts. They were pioneers, not polished.
The Original Pack (v1 — DALL-E era, 2022)
These were the originals, added December 31, 2022. Notice the inconsistent styles — some 3D cartoon, some semi-photorealistic:

magic

bike

lifting-weights

nerd

quill

family

meditate

vacation
These served the blog well for over two years, but the style was inconsistent across images and clearly “early AI generation.”
The Refresh (v2 — GPT Image, 2025)
In August 2025, the eulogy images got a full refresh using OpenAI’s gpt-image-1 model. This generation established the canonical raccoon brand:
- Rainbow round glasses
- Green t-shirt with bold white text (unique per image)
- Blue left Croc, yellow right Croc
- Soft plush 3D/vinyl illustration style
- Studio softbox lighting
- Subtle vintage film grain
- Children’s book aesthetic
The prompts were codified in scripts/generate_igor_images.py so images could be regenerated consistently. This was a huge leap — from “interesting AI experiment” to “coherent visual identity.”

DEALER OF WONDER

CAR-FREE SPIRIT

FIT FELLOW

TECHNOLOGIST

PAIN GOOD!

FAMILY

IGOR ♥ TORI

ON VACATION
The Brand System (v3 — 2026)
Through early 2026, new raccoon images were added for specific pages, all following the established brand:
The March 2026 Audit
In March 2026, we did a full audit of every raccoon image. Three images were off-brand and got replaced:
Updated references: 13 pages still pointed to old .png versions — all updated to .webp (including fixing a “vaccation” typo in 4 time-off posts).
Cleaned up: Deleted unused raccoon-ai-native-em.webp (duplicate, zero page references).
Final count after audit: 16 raccoon images, all on-brand.
The 7 Habits Marathon (May 2026)
In May 2026 I wired hero raccoons into all seven 7 Habits chapter posts. It took four iterations (v1 → v4, 21-35 cells per round, 7 habits × 5 variations on the final pass) and turned into the most useful pattern of the season: stop asking AI for “the best” raccoon — generate a wide field and let me pick.
Final picks, one per habit:
The choice-sheet pattern. The bottleneck on this batch wasn’t “can the model generate” — it was “which one do I want.” So I built a click-to-pick HTML sheet: 7 rows × 5 variations, localStorage-persistent picks, JSON copy button at the bottom, mobile-responsive. ~15 minutes of human picking and the right raccoons were obvious — much faster than iterating one prompt at a time. Live demo at /image-selector, reusable pattern published as a public gist (HTML/JS + README + CC0). The blog instance base64-inlines all 35 raccoons so the page is fully self-contained.
Three things I learned, distilled (full entry):
- Reference images beat prompt prose for character locking. Passing
raccoon-nerd.webpas a reference to Gemini Pro pulled canonical style 100% of the time. No amount ofIMPORTANT: rainbow glasses, mismatched Crocs, TECHNOLOGIST shirt…ever matched one reference image. Imagen ignored the style block entirely — a permanent note now says never Imagen for canonical raccoon. - Recraft’s
removeBackgroundbeat my chroma-key hill-climb. I’d been sharpening a local pipeline for six attempts; switching to a paid bg-remover at a penny per image solved it in an afternoon. The Wrong Jungle lesson from the day before applied immediately to this batch. - AI helps me think by showing me options. When the work is a judgment call I’ll live with, AI’s job is to widen the option space, not narrow it. I often can’t articulate my taste in the abstract — but I can tell you which raccoon I want when I see them side-by-side. Cheap model for the wide pass, human picks, expensive model only for the winners.
Final count after May 2026: 23 raccoon images, all on-brand.
Character Design Reference

The canonical raccoon character specification (from generate_igor_images.py):
Cute anthropomorphic raccoon character, big rainbow round glasses, green t-shirt with bold white text, blue left Croc and yellow right Croc, soft plush 3D/vinyl illustration, big friendly eyes, studio softbox lighting, transparent background, subtle vintage film grain, children’s book style. Full body.
Each image varies only the action/scene and shirt text. This constraint is what makes the brand work — instant recognition across 16 images.














