ip_illustration_for_yourself: a Mengli-style character anchor skill for Codex

What ip_illustration_for_yourself actually solves

Most image generation workflows break character identity the moment you change the task. You generate a cute portrait, then an article illustration, then a sticker sheet, and the face drifts each time. This project attacks that specific failure. It asks for a character reference image first, converts it into what the README calls a Mengli-style character anchor, and then requires every later asset to reuse that anchor. The anchor locks hair outline, fringe parting, side and back hair length, face shape, eyes, skin or fur colour, high-recognition features such as hats, glasses, ears and horns, body proportions, usual clothing, shoes, accessories, and the primary and secondary colours. The README states that one clear front-facing image plus one full-body image is the most reliable input. The audience is people who publish under their own name or brand and want a repeatable visual identity: writers who illustrate their own posts, solo creators, and small teams building a mascot. It is not aimed at someone who wants a single one-off picture.

How the anchor and the asset routes fit together

The architecture is a fan-out from one identity source, not a single generation pass. The README's diagram runs from a person, pet or character reference image, through the Mengli-style character anchor, into article illustrations, infographics, real-photo merges, stickers, folder icons, letter paper, Polaroid frames, scene avatars and emoji packs. Each route gets its own planning pass, so different aspect ratios and purposes are not squeezed into one generation. The README is explicit that the skill reuses the same anchor and plans each asset line separately. Two routes show how the constraints differ. Article illustrations are opinion-first: the full character plus props usually occupies about 20 to 35 percent of the frame, there is more white space, and text is normally omitted. Emoji packs are reaction-first: the face, hands and emotion must read at chat-thumbnail size, so the character and necessary props take about 45 to 70 percent, and 4 to 6 of the default 12 images get a partial, irregular small scene base such as a desk corner, grass, a bed, clouds or a window edge. The small scene must not fill the canvas or become a rectangular colour card. That is a real design decision: the two outputs share a style but not a composition rule.

Installing the skill and running a first anchor

There is no package to install. The README gives a single clone command and then tells you to place the whole repository into your Skills directory, or load it in a client that supports Skills. The repository layout matches that: SKILL.md sits at the top level next to README.md, README_EN.md, REFERENCE_URLS.md, and the agents, img2show, references and scripts directories.

git clone https://github.com/EverettFish/ip_illustration_for_yourself.git

After cloning, move the directory into your Skills folder and restart or reload the client so it picks the skill up. The README does not document a configuration file, an environment variable or a CLI entry point, so there is nothing to set beyond the directory location.

The first real use is a prompt, not a command. The README's suggested opening line is:

这是我的参考图，先建立一个萌粒风角色锚点，再用它做一套 All-in-one 个人 IP。

In English: this is my reference image, build a Mengli-style character anchor first, then use it for a full All-in-one personal IP set. What you should see is a proposed anchor, meaning a front-facing anchor image plus a three-view sheet, which you inspect for hair, face, clothing, accessories and proportions. The README is firm here: if the anchor is not confirmed, do not batch the rest. Once it is confirmed, you can request a single route, for example five article illustrations and two infographics, or three photo merges, or the full All-in-one set including stickers and folder icons.

The emoji takeover feature and its extra questions

The most distinctive route is the one the README calls emoji takeover. You upload a specific emoji or meme reference, and the skill replaces the complete identity of the original subject: head, body proportions, limbs, hands and feet, skin or fur, and species features. Before generating, it asks whether the whole set should follow the reference's original art style, the Mengli style, or a specified third style such as 3D or pixel art. Once chosen, the person, clothing, props, background, effects and text rendering are unified into that one style. There are two rules worth noting. First, the skill separates the body from what is genuinely worn: fur colour, belly, claws, ears, tail and animal torso must not be misread as clothing. If the original has real clothes, they are kept and fitted onto the IP body; if it has none, the anchor's default outfit is used. Second, composition follows the original at thumbnail-level geometry: character proportion, cropping, centre of gravity, tilt, limb angles, hand placement, foot line, text blocks and white-space distribution all track the reference. The repository ships three previously validated default references (silence/stop, chubby taunt, OK/hold on). If you have not uploaded anything, the skill asks whether to try some of those defaults; if you have uploaded references, it will not mix the defaults in automatically.

Where the skill breaks down

Real-photo merging is the route most likely to produce a rejected result, and the README says so indirectly by listing the failure modes. The final deliverable must be the merged image itself, not a transparent character layer pasted onto a photo. The character has to interact with a specific object or surface: leaning on a railing, sitting on steps, peeking from behind a display stand, holding a door handle, pointing at a real sign, or observing an exhibit. At least two spatial cues are required, such as foreground occlusion, a real contact surface, contact shadow, reflection, perspective match, or local light and colour match. The original photo's dimensions, crop, viewpoint and overall composition must stay unchanged, and non-interaction regions such as faces, products, text, buildings and exhibits must not be repainted. The default visible character height is 18 to 30 percent of the final photo height, with 22 to 26 percent preferred for ordinary scenes. A full character laid flat on the photo with no occlusion or contact, looking like a floating sticker, is explicitly judged unacceptable. That is a narrow target, and it depends on an image editing tool that can do local generation while leaving the rest of the frame intact. If your tool cannot, this route is the wrong one for you.

Transparency, folder icons and the Alpha check

The folder icon route is a good example of the project's verification stance. The default is four different schemes, not one icon recoloured four times: the character peeking from the top, sitting or leaning on the folder edge, emerging from a front pocket with a small action, and embedded in a miniature themed scene on the folder face. All four share a 4:3 folder master and the same character anchor, but change pose, peek position, outline direction, folder colour and one small themed prop. The README then states a delivery check that many pipelines skip: the file must have a real Alpha channel. Something that looks like a checkerboard but is actually an RGB image does not count as a transparent deliverable. The same logic applies to Polaroid frames. The README also gives macOS icon replacement steps: open the transparent PNG in Preview, press Command+A then Command+C, select the folder in Finder, press Command+I, click the small folder icon at the top left of the info window, then press Command+V. To restore the default, select the small icon again and press Delete. A Windows variant is offered only when the user explicitly asks, centring the 4:3 transparent image in a 1:1 transparent canvas and exporting .ico.

How it relates to plain image prompting

The obvious alternative is prompting an image model directly, either through a chat client or an API, with a description of your character repeated in every request. The difference is where identity lives. With direct prompting, identity lives in the prompt text and the model's short-term context, so it drifts as soon as you change the task, the aspect ratio or the scene. With this skill, identity lives in a separate anchor artifact that is confirmed once and then reused, and the skill layers per-route constraints on top: character height percentages for photo merges, 58 to 70 percent writable area and 9 to 12 light hand-drawn rules for letter paper, a 3:4 sheet with a light themed header card and 10 to 20 non-overlapping die-cut stickers below for sticker pages. Those numbers are not things you would naturally write into a prompt, and they are what makes the outputs look like one series. The trade-off is that you must accept the project's aesthetic. If you want a polished commercial illustration style, the README's own description of Mengli, with slightly shaky black pen lines, natural pauses and small irregular gaps in the outline, clean colour blocks with slight white showing at the edges, is not that.

Maintenance, licence and what is not documented

The repository is not archived, and the last push was on 2026-08-31. There are no retrieved releases, so there is no version history to upgrade against and no changelog to read before pulling. Upgrading means pulling the main branch and replacing your Skills directory copy, which also means any local edits you made to SKILL.md or the files under references and scripts are overwritten unless you track them yourself. The licence is not stated in the repository, and the README does not mention one. That matters for anyone planning to ship generated assets commercially: without a licence file, the terms under which you may redistribute or modify the repository are unclear, and the README does not address the copyright status of images you generate with it. Treat that as an open question to resolve with the author before commercial use. Also undocumented: rollback, any pinned image-model version, and any test suite. The README states that in Codex the skill uses the built-in image generation or editing tool when available and does not check, require or limit the underlying model's name or version, which means output quality will vary with whatever backend your client provides.

Editorial conclusion

Adopt it if you already work inside Codex with image generation available and you want one character to stay recognisable across illustrations, stickers, folder icons and photo merges. Do not adopt it if you need a standalone app, a hosted service, or a documented licence, because the repository ships none of those. Before you commit, verify that your Codex build exposes image generation or editing, and check whether the repository carries a LICENSE file, since the README does not state one.

Frequently asked questions

Official sources

1. EverettFish/ip_illustration_for_yourself on GitHub
2. Issues
3. README