Skill v1.0.0
currentAutomated scan100/100version: "1.0.0" name: aetox-design description: งานออกแบบภาพในแอตทอกซ์ - โลโก้ ไอคอน แบนเนอร์ ภาพโซเชียล และชุดอัตลักษณ์องค์กร (CIP) พร้อมแคตตาล็อกสไตล์ สีตามอุตสาหกรรม ขนาดมาตรฐานของทุกแพลตฟอร์ม และวิธีเขียนพรอมต์ภาพ อ่านตัวนี้ก่อนรับงานภาพเสมอ เพราะแอตทอกซ์ไม่ได้เจนรูปเอง มันหารูปจริงจากเน็ตมาใช้ หรือวาดเองด้วย SVG กับ HTML และสองทางนั้นเปลี่ยนวิธีรับงานตั้งแต่ประโยคแรก source: https://github.com/claudekit (design) license: MIT copyright: Copyright (c) claudekit contributors
Design work in Aetox
One fact decides everything below, so it goes first.
This app has no image model. Nothing here generates a picture from a description, not a logo, not an icon, not a photo. Saying otherwise, or quietly producing a prompt and calling it a deliverable, is the failure this document exists to prevent.
What it has instead are two hands that are real:
- Find a picture that already exists, on the web, and put it where the work
can use it.
- Draw it itself. SVG and HTML are text, and
writewrites text. A
wordmark, an icon set, a banner layout, a chart, a card, those are things to author, not to wish for.
So "make me a logo" is answered by choosing one of three roads and saying which one you took:
- Draw it, an SVG wordmark or lettermark, written here, editable forever,
and it prints at any size. This is the strongest answer for anything made of type and geometry, which is most marks.
- Find it, for photography, backgrounds, textures, and mockup scenes,
where a real photograph beats anything a description could produce.
- Spec it, when the user has an image model somewhere else (their own
account, a connected MCP server), the deliverable is a brief and a prompt sharp enough to spend money on. The style, colour and industry tables in this skill are exactly what makes that prompt good.
Never road 3 by default. It is the one that hands the work back to the user.
Getting a picture, or a sound
Three tools, in this order. This recipe is the canonical one; other skills point here rather than repeating it. It is the same recipe for music and sound effects, from the same kind of source (Pixabay Audio, Freesound, Wikimedia Commons carry sound the way Unsplash carries photos).
- `web_search`, find the page, not the file. Name the source in the query
when the licence matters: unsplash dark server room, wikimedia commons <subject>, pixabay whoosh sound effect.
- `web_fetch` on that page, it reads the HTML and lists every image URL it
found, with alt text. That listing is where the actual file URL comes from. Fetching an image URL directly is not useful: web_fetch reads bytes as text, so it returns garbage for a JPEG.
- `media_fetch` with that file URL and a destination path. It downloads,
checks the bytes really are an image or a sound (a saved HTML error page named .jpg is the ordinary way this goes wrong, and it refuses that with the page named), and reports the format, size, pixels or seconds in the same receipt — no shell, no separate verify step.
Licence, every time. A picture in a user's deck or brochure is a picture they are publishing. Prefer sources that say "free to use" on the page Unsplash, Pexels, Wikimedia Commons, a government or museum open collection and tell the user where each one came from. Never present a picture found by image search on an unknown page as cleared for use.
Third-party brand logos
A different case from either road above: an integration list, a "used by" wall, or a tech-stack badge needs someone else's mark, exact — not this app's approximation of it. Drawn from scratch (road 1) it drifts from the real thing; found by ordinary image search (road 2) it usually arrives as a raster screenshot, not the clean vector a logo bar needs.
[svgl.app](https://svgl.app) (github.com/pheralb/svgl, MIT) is a searchable library of pre-vectorized, web-optimized SVG logos for software, AI, and platform brands. Search it by name and use web_fetch + media_fetch on the entry's file under static/library/ the same way as any other found asset. A logo bar wants them as imgs/logo-*.svg beside the page, at one shared height, never at one shared width — the eye reads the cap height, not the bounding box.
The MIT licence covers svgl's own site and code, not the logos in it — each mark stays the trademark of the company it names; svgl's own contributor rules say as much ("ensure you have the right to use it and that its licence permits it"). Naming a real integration, dependency, or a real customer's logo with their consent is the ordinary use these marks exist for; implying endorsement, using a competitor's mark, or folding one into your own identity is not. When it's unclear, follow the entry's own brandUrl (svgl links it) to that company's brand guidelines. This is about sourcing other companies' marks — for rules on using this project's own logo, see aetox-brand/references/logo-usage-rules.md.
Drawing it instead
write produces the file directly, and for these that is the whole job:
- SVG, logos, icons, badges, diagrams, illustrations built from shapes.
Vector, so it scales, and it stays editable as text.
- HTML + CSS, banners, social cards, one-page layouts. To turn one into a
picture, build it as a single-slide deck and use the slides room's export bar, which writes .png, .jpg and .webp. Read aetox-slides before writing that file: the room has an anatomy, and a file that ignores it will not page.
คิดสองรอบก่อนวาด
ก่อนเขียน SVG หรือ HTML จริง ร่างแผนสั้นๆ ในหัวก่อน แล้วสอบแผนนั้นกับโจทย์อีกที, สองรอบนี้ถูก กว่าการแก้ทีหลังเสมอ
รอบแรก, ร่างแผน สี่ส่วนสั้นๆ พอ:
- สี, 4-6 hex ที่ตั้งชื่อได้ว่าแต่ละสีทำหน้าที่อะไร ไม่ใช่แค่บรรยายอารมณ์
- ตัวอักษร, อย่างน้อย 2 บทบาท (ตัวเด่นสำหรับหัวเรื่อง ใช้แต่น้อย + ตัวอ่านสำหรับเนื้อหา) เพิ่มตัวที่ 3
สำหรับตัวเลข/คำอธิบายถ้าจำเป็น
- โครงร่าง, แนวคิดเลย์เอาต์หนึ่งประโยค
- จุดจำ, หนึ่งองค์ประกอบที่อยากให้คนจำได้ว่าเป็นงานนี้
รอบสอง, สอบแผนกับโจทย์ ถามทีละส่วนว่า "ถ้าเป็นโจทย์อื่นที่คล้ายกัน จะออกมาแบบนี้เหมือนกันไหม" ถ้าใช่ แปลว่าส่วนนั้นเป็นค่าเริ่มต้นที่หยิบมาใช้เอง ไม่ใช่ทางเลือกที่มาจากโจทย์นี้จริงๆ แก้ก่อนลงมือ แผนที่ผ่านรอบนี้แล้วให้ยึดตามนั้นทั้งหมดตอนวาด ไม่ตัดสินใจสีหรือฟอนต์ใหม่กลางทาง
สามแบบที่มักหยิบมาใช้เองโดยไม่ทันคิด, พื้นครีมอุ่นกับตัวอักษร serif กับสีเน้นดินเผา, พื้นเกือบดำกับ สีเน้นเขียว/แดงสด, เส้นบางแบบหนังสือพิมพ์คอลัมน์แน่น, ไม่ได้ผิดในตัวเอง แค่ต้องมาจากการเลือกจริง ไม่ใช่ปฏิกิริยาสะท้อน ถ้าโจทย์หรือผู้ใช้ระบุทิศทางมาแล้ว ทำตามนั้นตรงๆ แม้จะตรงกับแบบใดแบบหนึ่ง ข้างบนพอดี คำสั่งที่ระบุมาชนะเสมอ
เช็กตัวเองรอบสุดท้ายแบบ Chanel: ก่อนเสร็จ ลองถอดหนึ่งองค์ประกอบออกดู ถ้างานยังสมบูรณ์โดยไม่มีมัน องค์ประกอบนั้นไม่ควรอยู่ตั้งแต่แรก
(แนวคิดจาก [anthropics/skills, frontend-design](https://github.com/anthropics/skills/tree/main/skills/frontend-design), Apache-2.0)
What a drawn asset owes, before it is handed over
The style is yours. The bar is not, and it is the same bar whichever model draws:
- A mark reads at 16 px and at 1024 px. Both rendered and looked at
(browser capture of the SVG, or the SVG placed in a one-slide page at both sizes). Detail that vanishes at 16 px is detail the mark does not have; a mark that only works large is a poster.
- On light and on dark, and a one-colour version, with
currentColor
or a named fill so the user can recolour it. Clear space stated as a fraction of the mark's height.
- An icon set is one set. One grid (24 or 20), one stroke width, one
corner radius, optical alignment checked side by side in one row, not each icon in its own file. Two icons that look drawn by two hands are.
- A banner's text sits inside the platform's safe area from
references/banner-sizes-and-styles.md, at the size that platform crops to, and the file is the exact pixel size the table gives.
- Type is real type. A wordmark set in a font names the font and its
licence, or is converted to paths so the SVG does not fall back on a machine without it.
- Named as the user will find it, and the SVG is tidy: no editor
metadata, no unused <defs>, ids that say what they are. It is a source file, not an export.
What is here
Knowledge, not commands. Open one with skill_view and a path.
| Question | File | |
|---|---|---|
| Which mark style suits this brand | references/logo-style-guide.md | |
| What a colour will say about them | references/logo-color-psychology.md | |
| How to write an image prompt that is worth spending on | references/logo-prompt-engineering.md | |
| What a corporate identity programme contains | references/cip-deliverable-guide.md | |
| How those pieces should look together | references/cip-style-guide.md | |
| Prompting for identity mockups | references/cip-prompt-engineering.md | |
| Every banner size, per platform, and the styles that work | references/banner-sizes-and-styles.md | |
| Sizes and craft for social photos | references/social-photos-design.md |
The tables are the catalogues, read whole rather than searched, each is small, and reading it whole is how you see the row you would not have thought to look for.
| Table | Rows about | |
|---|---|---|
data/logo/styles.csv | mark styles | |
data/logo/colors.csv | palettes and what they carry | |
data/logo/industries.csv | conventions per industry | |
data/icon/styles.csv | icon styles, usable directly when drawing SVG | |
data/cip/deliverables.csv | what a full identity programme ships | |
data/cip/styles.csv | identity styles | |
data/cip/industries.csv | industry conventions | |
data/cip/mockup-contexts.csv | scenes to show an identity in |
Where the neighbouring work lives
- Decks and presentations,
aetox-slides. It owns the anatomy of a deck
this app can page through and export; this skill does not.
- **Tokens, component specs, and the tables that decide a slide's layout,
typography and charts**, aetox-design-system.
- Voice, messaging, logo usage rules, approval checklists,
aetox-brand.
One accent colour doing every job, three weights of light rather than pure white on pure black, and the same furniture on every surface: that is the house look, and it is a reference rather than a rule.