Illustration Extract Template logo

Illustration Extract Template

Organization
okou-ai
illustration-extract-template

Extract a style prompt from a reference picture, show the user pictures made from it, revise until they approve, then save it as their template. Use when a user uploads an image style to reuse, asks to extract a prompt from a picture, wants their own image template, asks to train, forge or lock a new illustration style, or asks for more images in the style of a picture they supplied.

Overview

Publisherokou-ai
Repositoryokou-skills
Skill nameillustration-extract-template
Stars
81
Forks
19
Bundled files
3
Links
  • Markdown instructions

    A SKILL.md file the model loads on demand, so it only costs tokens when a request actually matches.

  • Works with any LLM

    AI skills are plain Markdown, not provider-specific code, so this works with GPT, Claude, Gemini, Grok, or a local model.

  • 3 bundled files

    Scripts, templates, and references the model can read while it works. Files are read-only and never executed.

  • Open source

    Published by okou-ai on GitHub. Read the source before you install it.

Installation

Install the Illustration Extract Template AI skill in TypingMind to use it with any LLM, or drop it into another agent that reads SKILL.md.

1

Install in TypingMind

TypingMind installs a skill straight from its GitHub folder — it reads SKILL.md, bundles the resource files, and stores the result locally.

  1. Open the app and go to Plugins → Skills.
  2. Choose "Install from GitHub".
  3. Paste the skill folder URL below and confirm.
    https://github.com/okou-ai/okou-skills/tree/main/extract-template/illustration
  4. Enable the skill in any chat where you want it available.
Plugins → Skills → Add skill → From GitHub URL, then paste the folder URL and press Continue.
2

Install in another agent

Any agent that reads the Agent Skills format can use this skill — copy the folder into that agent's skills directory.

Claude Code — .claude/skills
git clone --depth 1 https://github.com/okou-ai/okou-skills.git /tmp/okou-skills
mkdir -p .claude/skills
cp -r /tmp/okou-skills/extract-template/illustration .claude/skills/illustration-extract-template
Restart Claude Code after copying so it picks up the new skill.

Use it in TypingMind

Enable Illustration Extract Template in any TypingMind chat and the model takes it from there. Its name and description sit in the system prompt, and the moment a request matches, the model loads the full instructions itself — you never invoke it by hand, and it costs no tokens until it is actually used.

The model loads Illustration Extract Template on its own as soon as a request matches it.

Works with any AI model

AI skills are plain Markdown instructions rather than provider-specific code, so Illustration Extract Template is not tied to the model it was written for. Install it once in TypingMind and use it with GPT-5, Claude, Gemini, Grok, DeepSeek, Mistral, Llama, or a local model you run yourself — all on your own API keys.

  • Loaded only when it is needed

    The system prompt carries just the name and description. The instructions are fetched on the first matching request, so an idle skill costs nothing.

  • Switch models mid-chat

    Because the skill is instructions rather than code, changing model does not break it — the next model reads the same SKILL.md.

Skill instructions

This is the SKILL.md content the model loads. Read it before installing — a skill is instructions your model will follow.

Extract a reusable style from a reference picture

Input: images of one style, or a brief when no reference exists. Output: the user's own style, saved once they have seen what it makes and said it is right.

No extraction is complete. Measurement settles canvas, ground, palette, stroke, coverage and placement. Medium, drawing and subject convention are read by eye and are partly wrong on the first pass. The user closes that gap by looking at pictures, so the path is: extract, show, revise, save.

Choose among your own candidates without asking — the gate ranks them. Stop before saving, always: never save a prompt the user has not seen pictures from.

Run every command below from extract-template/illustration/.

Before anything — a style, or this one picture?

A style makes new pictures: another subject, another scene, the same look. A picture is one arrangement the user wants back.

For one picture, write the prompt that recreates it and stop. Continue here only for a style.

Prerequisites

bash
python3 -c "import pymupdf"     # else: pip install pymupdf

Steps

1. Collect the references

Download every [Web file] with okou web download-file.

The scripts read PNG, JPEG, TIFF, BMP, GIF, JPEG 2000 and Netpbm. They do not read WebP, AVIF, HEIC or ICO — the formats a phone or a web page hands over. Convert those before step 2, and keep the original:

bash
ffmpeg -loglevel error -y -i <ref.webp> <ref.png>

Ask for what is missing, and nothing else:

  • intended use — blog cover, product card, marketing poster, in-app spot illustration. It decides the canvas and the frame.
  • a slug — kebab-case, the name the saved style is called by. Propose two or three from the reference if the user has none.

With no reference at all, write the frame from the brief and go to step 4. Steps 2 and 3 need pixels.

Three or more references separate a locked axis from a dial. With one, ask for more; if none arrive, record in design-system.md which axes stay unsettled and give them no range.

Ask for files at least 400px on the short side. Below that the palette fills with anti-alias blends and the stroke is one pixel wide, so colour and line describe the downscaling.

2. Measure

bash
python3 scripts/measure_style.py <ref> [<ref> ...] --json style.json

Take these numbers over your own reading of the image. The verdict column decides where each axis goes:

VerdictWhere it goes
CONSTANTthe locked frame, as a value
RANGEthe locked frame, as a range
VARIESa dial

The shared-colour line is the locked palette; every other colour is a palette dial.

3. Read what the numbers cannot reach

Keep two records apart:

  • content — subject, action, props, setting, placement;
  • style — how it is drawn, carrying no content noun.

Then delete every subject, object, place and name from the style record. What remains must still describe a way of drawing; move anything else back to content.

Write every style line as a drawing instruction that names the near miss it excludes. A line that only names the family collapses into the nearest common style — flat vector, cel anime, stock watercolour — whatever the reference actually was.

Instead ofWrite
flat gouache shapespaint varies in value inside every shape, never a flat digital fill
thin dark contourshapes meet at their colour edges, no outline; only the eyes and leaf veins carry a drawn line
botanical motifseach sprig painted with its own texture and veins, not a single-colour silhouette
soft watercolourwet washes that bloom and granulate; almost nothing in the picture is a flat area
closed eyesheavy lids as grey-shadowed lens shapes under one fine dark line

State the style under these heads, from the references only:

HeadState
Mediumpainted, inked, vector, 3D, collage, print, photographic, mixed
Linethe measurement's contour verdict first, then closed or open, uniform or tapered, cap shape
Shaperounded or angular, geometric or organic, how forms simplify
Valueflat fill, cel steps and how many, gradient modelling, hatching, wash
Colourwhat the lead, support and accent each do
Texturebrush, grain, paper, halftone, noise — and where it sits
Detailwhich areas carry detail, which stay plain
Lightdirection, rim, glow, haze, or no light description at all
Compositioncrop, weight, repeated shapes, the rhythm every reference keeps
Framehow the art meets the canvas: edge to edge, a vignette, or a visible margin — and how wide
Finishgrade, bloom, chromatic shift, fade, print registration
Subject conventionface treatment, cast, count, scale, what the subject does

Then name the dials — what changes from piece to piece. Three to seven. These carry across styles:

DialWhat it varies
Palettea hex set, or a named palette family
Scene metaphorthe per-piece concept. The scene IS the metaphor — never a generic filing cabinet with relabelled tabs
ComplexityL1 single subject / L2 small scene / L3 full vignette
Castthe character in this piece, not a mascot the style is stuck with
Propsnative to the theme — a crane for construction, a press for publishing
Accent marksthe small repeated motif: dots, ticks, sparkles, ink dabs
Moodposture and register, built from what is visible

4. Write the style prompt

Write the prompt first; it is the deliverable. Its first third carries the locked frame, then the subject, then the dials, then what must not appear.

Keep these behind it, and write them into the package at step 8:

  • Locked frame — one line per axis, measured values included. State the contour width in pixels at the canvas the package delivers, not only as a percentage: the references are rarely that size. State how much of the sheet the art covers and whether the ground stays unpainted — without it the model floods the sheet and the paper stops being a colour.
  • Dials — one line per dial named in step 3, with the values the references used and what else the dial may take.
  • Not in the frame — techniques absent from every reference.
  • Prompt template — one prompt with a placeholder per dial, opening on the locked frame.
  • References — one line per file naming the dial values it demonstrates.

design-system.md carries the measurement table per reference, the verdict table, and the axes a single reference could not settle.

Write the prompt template so the locked frame occupies its first third.

5. Generate, measure, choose

bash
npx --yes --package="${CLI_PKG_URL}" okou generate image --provider built-in \
  --raw-prompt "<the style prompt, dials filled in>"
python3 scripts/check_piece.py --refs <ref> [<ref> ...] --piece <generated> [...]

Generate three at a time — that is the ceiling on generations in flight — each on a subject the references do not carry, and each moving a different dial: another scene metaphor, another complexity level, another palette value. Three pieces of the same subject show the style sitting still, not flexing.

Keep the ones with the fewest axes outside range. Use the built-in image model; if it fails for this style, say so and ask how to proceed rather than routing to another provider.

Read the failing axes: they name what the prompt did not hold. Correct the prompt and generate again.

Some axes never come back from the prompt. When two rounds do not move one, stop rewriting and fix it after the fact where that is possible. Scale and placement on a plain ground is the common one:

bash
python3 scripts/compose.py --ref <reference> --piece <generated> --out <placed>

It scales the drawing's ink box to the fractions the reference measures and places it at the reference's margins, touching nothing inside the drawing.

Show only pieces the gate passes. When an axis will not come back, keep the best set and carry that one sentence into step 6.

The check cannot see medium, line quality, shape language, subject conventions, motif or composition. Look at the kept piece for those before delivering.

6. Show the pictures and ask whether the style is right

Put in the reply:

  • the pictures as markdown images, one line under each naming its subject and the dials it moved — "L1, sage palette, single character";
  • one line saying these were generated from the style in their reference.

Then ask one question: does this style look right.

This is a yes or no, not a survey. Do not list the axes, do not ask what to fix, do not offer options. When something still departs from the reference after step 5 did what it could, say it in one plain sentence — the lines came out heavier than the reference, the palette runs warmer — and leave it there.

Show the prompt only if the user asks for it.

Stop. Save nothing yet.

7. Revise, then show again

The user speaks about the look — the lines are too heavy, it is too colourful, the faces are wrong. Turn each correction into one line of the locked frame that names the near miss it excludes, the same form as step 3. Nothing else in the prompt changes.

Regenerate the set with step 5 and show it again. One round of corrections at a time; do not bundle two rounds into one reply.

Repeat until the user says it is right.

8. Save it as the user's template

Only after the user says the style is right.

Build the package under the slug from step 1:

text
<slug>/
  SKILL.md                 frontmatter with the trigger phrases, the locked
                           frame, the dials, the prompt with its placeholders,
                           and two or three example briefs
  design-system.md         the measurement table per reference, the verdict
                           table, and the axes a single reference could not
                           settle
  ref-<subject>-<dial>.png the approved pictures

Leave the user's references out of the package; the primary one is published beside it as the source. With no reference at all, publish an approved picture as the source instead.

Generate one piece from the SKILL.md you just wrote rather than from the prompt in your notes, and check it still matches the approved pictures. Publish only after that.

bash
npx --yes --package="${CLI_PKG_URL}" okou user-template publish \
  --title "<user-visible style name>" \
  --kind illustration \
  --source <the primary reference> \
  --package <slug>

--source takes PNG, JPEG, BMP or WebP. Convert anything else — AVIF, HEIC, GIF, TIFF, JPEG 2000, Netpbm — and pass the PNG:

bash
ffmpeg -loglevel error -y -i <ref.heic> <ref.png>

Say the template exists only after the command succeeds; name the specific blocker if it fails.

Then report, in this order: what the style is called and that it is under Custom in the template picker, the link to the picture the user approved, and anything still open — a dial with one value, a complexity level never tested, an axis that departs from the reference.

Register it as a built-in selectable style only when the user asks: the resource goes to illustration-template/<slug>/ in okou-ai/okou-skills, its entry to the Open Design registry in okou-ai/okou as vm0:image-style:<slug> with a selection description of 150 characters or less, and each pull request links the other.

Rules

  • Reproduce the references. Do not correct a palette, a proportion or a stroke you would have drawn differently.
  • Present in every reference is locked; present in one is a dial; with a single reference, neither — say so rather than guessing.
  • A technique absent from every reference is part of the locked frame. A subject or content type absent from them is not a rule.
  • Name observable technique, never an artist, studio, brand or product.
  • Numbers in the package come from step 2, not from reading the image.
  • Step 5 is mandatory. Show pictures you generated, never only a prompt.
  • Choose among your own candidates yourself. Do not hand the user a menu of variations.
  • Step 6 stops, and asks one yes-or-no question. Saving before the user has seen pictures and approved them is the one thing this skill must never do.
  • A correction from the user is one more line in the locked frame, not a rewrite of the prompt.
  • Rotate the cast and the scene across a series. A locked frame is not a locked mascot.
  • Use the built-in image model. When it cannot hold a style, say so; do not route to another provider without asking.
  • Nothing is published from an untested style. Generate one piece from the written SKILL.md first.
  • The user's reference files are never part of the package. The primary one is the published source; the rest served steps 2 and 3 and stop there.

Troubleshooting

SymptomAction
Generated pieces repeat the reference's subjectThe subject leaked into the style record; run the deletion test in step 3 again and rewrite the prompt template
check_piece.py fails on line widthStep 2's measurement, not your estimate, goes in the package
Generated pieces carry outlines the reference has none ofThe measurement said NO drawn contour; remove every outline word from the prompt and say shapes meet at their colour edges
A generated piece mounts the art inside a paper borderThe Frame line is missing or too weak; say the paint reaches all four edges
Generated pieces read as flat vector against a painted referenceFlatness under about 60% means paint varies inside each shape; say so, and name what must not be flat
Every generated piece looks the sameToo few dials, or the prompt template has no placeholders
A colour appears that no reference usesThe palette dial has no list of allowed values
Generated pieces flood the whole canvasThe locked frame is missing the ink coverage and the unpainted ground; "vignette" alone does not hold
The medium drifts to pencil or crayonName the wet behaviour — washes pooling at the stroke edges — and name the media to avoid
A generated piece renders on a black groundIt is RGBA with a transparent ground; the scripts composite over white, and so must anything you hand the user
The drawing fills the frame however the prompt words itTwo rounds is enough; place it with compose.py
The check passes but the piece looks wrongMedium, shape language or subject convention is missing from the locked frame
LOW RES in the measurementAsk for a larger file. Keep aspect, ground colour, ink coverage and centring; leave colour and stroke out of the locked frame
The reference is a scan or a photograph of a paintingIts hairline border is stepped past, and the ground axis is skipped when the art covers the canvas. Nothing else changes
The generator has no canvas at the reference's ratioGenerate at the nearest size and crop centrally to the reference's ratio
unknown image file format on a referenceIt is WebP, AVIF, HEIC or ICO. Convert to PNG with ffmpeg and measure that
Background reads as textured on a flat styleThe reference is a JPEG; re-export as PNG or accept the grain figure it reports
One reference onlyRecord the unsettled axes; do not write ranges you cannot support
The user approves without comment on the first showingSave it. Do not invite more rounds
The user asks what could be betterNow name the axes that still fail, two or three, shortest first
The user's correction contradicts the referenceFollow the user. Note in the saved file which axis now departs from the reference
Two corrections arrive at onceFold both into the locked frame, regenerate once, show once

Bundled files

The model reads these on demand while the skill is loaded. They are exposed as readable files and are never executed.

Frequently asked questions

What does the Illustration Extract Template AI skill do?

Extract a style prompt from a reference picture, show the user pictures made from it, revise until they approve, then save it as their template. Use when a user uploads an image style to reuse, asks to extract a prompt from a picture, wants their own image template, asks to train, forge or lock a new illustration style, or asks for more images in the style of a picture they supplied.

Why use Illustration Extract Template on TypingMind?

Because you install it once and use it with any model. Illustration Extract Template is plain Markdown rather than provider-specific code, so the same skill runs on GPT-5, Claude, Gemini, Grok, or a local model — and you can switch model mid-chat without it breaking. TypingMind runs on your own API keys, so you pay providers directly instead of a per-seat subscription, and your skills and chats stay in your own storage.

How do I install Illustration Extract Template in TypingMind?

Open Plugins → Skills → Install from GitHub in TypingMind and paste https://github.com/okou-ai/okou-skills/tree/main/extract-template/illustration. TypingMind reads its SKILL.md and bundles its files and installs it as a skill you can enable per chat.

Which AI models can use Illustration Extract Template?

Any model you connect in TypingMind. AI skills are plain Markdown instructions rather than provider-specific code, so GPT, Claude, Gemini, Grok, and local models can all load this skill when a request matches it.

How many AI models can I use with Illustration Extract Template?

As many as you like. As long as a model supports skills, you can use Illustration Extract Template with it — GPT, Claude, Gemini, Grok, DeepSeek, Mistral, Llama and more — all on TypingMind with your own API keys.

Is the Illustration Extract Template AI skill free?

It is published on GitHub by okou-ai. Check the repository for licensing terms. You only pay your own AI provider for the tokens you use.

What are AI skills?

An AI skill is a reusable instruction bundle that teaches an AI model how to do one specific task. It follows the open Agent Skills format: a SKILL.md file with a name and description, plus any scripts, templates or reference files the model may need. The model reads the instructions only when your request matches the skill, so an installed skill costs nothing until it is used.

How are AI skills different from plugins or MCP servers?

A plugin or MCP server gives a model new tools to call — code that runs somewhere and returns a result. An AI skill gives the model knowledge and process instead: how to approach a task, which steps to follow, what good output looks like. Skills are plain Markdown, so they need no server, no API key and no runtime, and they work with any model.

View all

Set up your own AI workspace now

Get notified about new features and future giveaways by subscribing to our newsletter 👇