Building a Reusable Material Illustration Skill with Codex+GPT-Image

歸藏的AI工具箱 · wechat · 2026-07-08

The author created a custom social media Skill allowing Codex to leverage GPT-Image 2.0 for generating 3:4 article visuals. The results are outstanding (clean white backgrounds, restrained 3D materials, and Chinese labels embedded directly). This led to the open-source Skill "guizang-material-illustration". It solves the challenge of illustrating concepts, processes, and data clearly—transforming articles, notes, or product manuals into Chinese explanatory diagrams with arrows, annotations, and spatial relationships for quick comprehension. The post details the engineering process from ad-hoc prompts to a stable Skill: designing visual structures and templates for scenarios like business reports, data charts, and educational illustrations, unified by a white-background studio light, restrained 3D materials, and IKB blue accents. For niche concepts, the Agent first retrieves references to standardize the style. The author repeatedly corrects the model's anti-pattern of "not writing text in images," requiring 2–5 character labels in clean areas. Instead of reskinning screenshots, charts are semantically extracted and redrawn from scratch, with a pre-delivery review to prevent pure atmospheric imagery.

Original post →

Building a Reusable Material Illustration Skill with Codex+GPT-Image

More from coding & agent