illo — Editorial-Illustration Agent Skill with a Recurring Mascot

Source: illo-skill-tmchow-2026-06-14 (GitHub repo + README + SKILL.md, fetched 2026-06-14), flagged as “ELO” in raw/GitHub_Trending_Weekly_36_-_ghostty-blackhole_LiteDoc_MiMo-Code_Bernini_UniRL_MSA_concord.md

illo (Trevin Chow, MIT, 90★ — the GitHub-Trending auto-caption heard the name “illo” as “ELO”) is an AI agent skill that turns an idea or a whole article into an original print-style editorial illustration starring a recurring mascot, so a blog’s images read as one consistent house style instead of cheesy stock photos or generic AI slop. It is deliberately not a generic image generator: “the methodology is the constant; the character pack and palette are the parameters.” The standout practical detail for this stack is its dual rendering backend — your Codex CLI (gpt-image-2, free on a Codex subscription) when present, OpenRouter as the universal fallback — and that it installs across ~70 agent runtimes from one repo.

Key Takeaways

• Solves the blog-header problem with a house style. One image explains one idea (a key judgment, a flow, a before/after, a trap, a loop); a recurring mascot performs the idea in every scene as the subject, never decoration. The default mascot is Blot, a deadpan ink-drop. The result is brand-consistent editorial art, the kind you’d see in a magazine.
• Ten bundled “looks” + a character builder. riso (house default — grainy halftone, ink-layer offset, paper grain), blueprint, woodcut, pixel, clay, manila, chalk, phosphor, enamel, gouache — or a custom style file. A character pack carries its style with it (one look per pack). Build your own mascot via the guided character builder; install community characters from tmchow/illo-characters.
• Two registers, not just one picture. An editorial scene (default) or an explainer — a hand-built flow / fan-out / timeline / loop / stack drawn in the active look with the mascot as a working part of the diagram — plus mini-comics (2–4 panels in one image) when an idea advances through stages. It’s intentionally not a photo, logo, corporate infographic, formal flowchart, or UI mockup.
• Dual backend = a real cost lever. The engine (scripts/illo.py, stdlib Python, no installs) renders through either: Codex CLI (gpt-image-2 on the user’s Codex subscription — no per-image charge, no API key, just shells out to the user’s logged-in CLI; macOS/Linux only, not Windows/WSL); or OpenRouter (model-selectable: Grok Imagine, Nano Banana 2/Pro, GPT-5.4 Image 2, …) as the universal fallback on any host without Codex.
• Disciplined trigger + key hygiene. The skill fires only when directly invoked or “illo” is requested — never on a generic “draw / illustrate / make an image” request — so it won’t hijack unrelated prompts. Secrets go through one channel: a user-run init writes a mode-600 config; the engine never reads keys from the environment or accepts them as CLI args, and the skill instructs the agent never to enter the user’s key itself.
• Brand-distribution as a feature. Installs across Claude Code, Cursor, Codex, Copilot, Gemini CLI, Hermes, OpenClaw, and ~70 other runtimes via the skills.sh CLI (npx skills add tmchow/illo-skill --skill illo), and ships native plugin/extension manifests per platform for managed updates. It follows the canonical skill-repo layout (anthropics/skills, openai/skills): a top-level skills/ folder, one dir per skill.

What makes it “not AI slop”

illo’s pitch sits squarely in the wiki’s anti-generic-AI-art thread (see Refero Styles and Leila Hormozi’s “I am so sick of reading AI slop” memo in How Alex Hormozi Uses AI). The mechanism is constraint:

• A fixed methodology (one metaphor per image, mascot-as-subject, the “load-bearing test” that the character must carry the idea) plus parameterized character + palette — palettes come from presets, your own file, or a derive-from-one-dominant-color algorithm so art matches an existing brand.
• Reference-image character consistency — a canonical model sheet (assets/character-reference.webp) anchors the mascot across renders, the same “pass the actual reference, demand exact fidelity” discipline documented in Higgsfield × Claude creative-agency work.
• An explicit “never copy their compositions — invent a fresh metaphor” rule and a post-generation quality-bar checklist.

Implementation

Tool/Service: illo (tmchow/illo-skill), v0.22.1 — Python agent skill, MIT. Setup:

# any of ~70 agents (recommended):
npx skills add tmchow/illo-skill --skill illo
# Claude Code native plugin:
/plugin marketplace add tmchow/illo-skill
/plugin install illo@illo-skill
# user (not the agent) bootstraps the key + non-secret defaults:
python3 "$SKILL_DIR/scripts/illo.py" init        # hidden prompt → ~/.config/illo/config.yaml (mode 600)
python3 "$SKILL_DIR/scripts/illo.py" doctor       # preflight; run standalone, never chained with &&

Cost: the skill is free (MIT). Rendering cost depends on the backend — **$0/imageviatheCodexCLI\ast \ast ifyoualreadypayforaCodexsubscription(drawsonyourCodexquota),otherwisemeteredthroughyour\ast \ast OpenRouter\ast \ast keyperthechosenimagemodel.\ast \ast Integrationnotes:\ast \ast invokewithanideaoranarticleURL(‘illo[ideaorarticleURL]|buildacharacter|install<character>‘).Foranarticleitroutesthesource\to thesis\to coverage(hero/hero+set/set/mini-comic),buildsashotlist,andrendersoneimageperanchor.Social-readyoutputis16:9(or1:1)withabold‘ink-punch‘andanoptional‘x‘-handlewatermark.Fixanexistingimagebypassingitbackas‘--ref‘.\ast \ast Hermes-onlycaveat:\ast \ast someHermesversionscorruptbundledbinarycharactersheetsoninstall(silentlybreakingthecharacterlock)—run‘bash${HERMES_SKILL_DIR}/scripts/repair-hermes-assets.sh` once; faithful runtimes (Claude Code / Codex / OpenClaw) don’t need it.

Ecosystem

A small cluster of editorial-illustration skills landed in the same window — useful to know before standardizing on one:

• orange2ai/orange-line-illustration (135★) — “New Yorker-style minimalist editorial illustration… one idea, one accent, lots of silence.” The most-starred direct competitor; note its license is non-standard (free for OSS use, commercial license required for closed-source) — unlike illo’s plain MIT.
• tmchow/illo-characters (10★) — community mascot packs for illo.
• KarenSpinner/article-thumbnail-skill — the same brand-block + recurring-character idea, but rendered via Gemini 2.5 Flash Image.

For WEO-style branded collateral the team already generates gpt-image-2 art through Codex (the codex-imagegen path); illo’s Codex backend is the same gpt-image-2-via-Codex pipeline wrapped as a reusable, mascot-consistent house style — worth a look as a structured alternative to one-off prompt files. ^[inferred connection to existing workflow]

Try It

• Install it into Claude Code (npx skills add tmchow/illo-skill --skill illo) and hand it one of your existing blog posts by URL — judge whether the riso default reads as “branded” vs your current stock-photo headers.
• If you have a Codex subscription, confirm the free path: run doctor, verify it detects the Codex CLI, and generate without an OpenRouter key to get $0/image art.
• Build a WEO/Smile-Springs mascot with the character builder and a palette derived from the brand’s dominant color, then generate a hero + explainer for one article to test cross-image consistency.
• A/B it against orange-line-illustration on the same article thesis to feel the maximalist-mascot vs minimalist-line tradeoff before standardizing.
• Mind the license if you ever wrap a competitor: illo is MIT (safe for closed/commercial); orange-line needs a commercial license for closed-source use.

Related

• agent-skills (Addy Osmani) — the skill spec-to-ship lifecycle and the skills.sh ecosystem illo distributes through.
• Refero Styles — same anti-”generic AI slop” thesis via brand-DNA constraints, applied to UI rather than illustration.
• How Alex Hormozi Uses AI — Leila Hormozi’s anti-AI-slop memo; illo’s house-style discipline is the visual-creative version.
• Higgsfield × Claude Creative Agency — reference-image fidelity / brand consistency in agent-driven image generation.
• GPT Image 2 Launch — the model illo renders with through the Codex backend.
• Lead Magnet Creation with Claude Code — where branded article art slots into a content/lead-magnet pipeline.

Open Questions

• gpt-image-2 vs the OpenRouter models — no head-to-head quality/cost comparison in the source for editorial-illustration output specifically; which backend gives the best mascot consistency is untested here.
• How well does character consistency actually hold across dozens of images from one model sheet? The claim is reference-anchored, but drift is the usual failure mode of recurring-character image gen — verify on a real run before promising clients a “consistent brand.”
• Codex-backend availability is macOS/Linux only and depends on the image_generation feature being enabled on the user’s Codex CLI — a moving target as Codex changes.