Storytelling Figures — Development skill for Claude Code

storytelling-figures

**v6.0** · **[English](#english) · [中文](#中文)**

A **Claude Code / Codex skill** that turns a conversation into a clean, paste-ready **image-generation prompt** for scientific, conceptual and explanatory figures — and quietly keeps the model from making the usual weird mistakes.


---

English

You're chatting with an AI about some mechanism — how an LLM tokenizes a typo, how an agent picks a skill, your own method's pipeline. At some point you think *"this is hard to explain, let me just draw it."* This skill catches that moment: it reads the conversation you already had, engineers a single high-quality prompt for **GPT image-2 / nano-banana**, and hands it back ready to paste.

It is **not** a figure renderer and **not** a chart library. Its whole job is to write *one good prompt* and avoid the failure modes that make AI-generated figures look wrong or tacky.

Highlights

A model can already write an image prompt. The catch is **aiming it** — the things you *can't* get by prompting an image-generation AI yourself, no matter how you phrase it:

What you get	By yourself	With the skill
🎯 Hits your target	❌ a vague prompt → a random draw, rarely the figure you pictured	✅ pins the few choices that narrow it to your figure — which points (≤5), how full, what to emphasize
🔒 No made-up / stray content	❌ invents steps, slips a passing background word in as a concept	✅ auto-builds a word allow-list (source-lock) — no invented, off-vocabulary terms
📐 Logically consistent	❌ lines from nowhere, content warped to make the layout look nice	✅ input↔output closure, no dangling lines, content not bent for looks
🎨 Reads without labels	❌ leans on text labels — strip them and it's unrecognizable boxes	✅ each concept drawn as a real object, legible with the text removed

![By yourself v