| name | edu-image-prompt |
| description | Generates structured 5-block image-generation prompts for educational visuals from any knowledge point. Produces copy-paste-ready prompts optimized for GPT Image 2, Nano Banana, and Midjourney with reliable Chinese text rendering. Use when the user wants to turn a concept, lesson, or topic into a knowledge card (图文卡片), infographic (知识地图), process diagram (流程图), classroom illustration (课堂配图), or social-media study card (知识卡片), or asks for a 生图 Prompt / AI 配图 / 教学素材. |
edu-image-prompt
Turns any educational knowledge point into a copy-paste-ready image-generation prompt that produces clean, readable educational visuals — not decorative posters.
When to use
- Turning a concept / lesson point / outline into a visual prompt
- Creating PPT slides, knowledge cards, process diagrams, lecture handouts
- Concept visualization for abstract theories or technical terms
- Social-media education cards (Xiaohongshu, public-account posts)
When NOT to use
- User wants the image itself generated directly (not a prompt)
- User wants a photo edit or image retouching
- User wants a logo, brand asset, or marketing creative
Workflow (3 steps)
Step 1 — Style setup
If the user already specified a style ("用黏土风", "做小红书卡片"), skip to Step 2.
Otherwise, use sensible defaults and proceed — do NOT block on a menu:
| Parameter | Default | Full menu (only if user asks "有哪些选项") |
|---|
| 领域美学 | 科技/AI 玻璃态 | A)科技 B)商学黑金 C)人文复古 D)少儿黏土 |
| 质感渲染 | 扁平 UI | A)等距3D B)扁平UI C)场景实拍+AR |
| 展示平台 | 通用讲义 16:9 | A)PPT/直播 B)小红书3:4 C)通用16:9 |
| 目标模型 | GPT Image 2 | A)GPT Image 2 B)Nano Banana C)Midjourney D)不指定 |
Rationale: most users just want a good card fast. Give them the default, let them override.
Step 2 — Knowledge decomposition (CRITICAL)
Fill this template. Do NOT skip — this is the step that separates a useful educational image from a pretty-but-empty poster.
Decomposition table (copy and fill):
| # | Visual goal (must be drawable) | Module | On-image text (≤8 chars) |
|---|
| 1 | [具体可绘制元素, e.g. "一句话被拆成积木块"] | analogy / definition / ... | "≤8字" |
| 2 | ... | ... | ... |
| 3 | ... | ... | ... |
| 4 | (optional) ... | ... | ... |
Module pool (pick 5-8, see type-guide.md for details):
definition · example · steps · comparison · pitfall · analogy · gold-sentence · rating · flow-arrow · zoom-detail · warning · mnemonic
Validation checkpoint — before proceeding, verify:
Step 3 — Generate, validate, and deliver
- Assemble the 5-block prompt using the template in
prompt-rules.md, applying the style from Step 1 and modules from Step 2.
- Apply model-specific techniques from
model-guide.md (e.g., GPT Image 2 thinking mode, MJ --ar parameter).
- Quick-check against the validation rules below.
- Deliver following the output format (prompt first, design notes after).
The 5-block prompt structure
Every prompt must follow this order (see prompt-rules.md for full template):
[1] CANVAS — dimensions, background, dominant accent color
[2] HEADER — exact title text wrapped in THE TEXT READS: "..."
[3] CONTENT BLOCKS — 5-8 named modules with labels
[4] VISUAL — line weight, icon style, arrow style, palette
[5] CONSTRAINTS — text accuracy rules + negative keywords
The #1 text-rendering rule: wrap every Chinese string in THE TEXT READS: "...". This is the biggest accuracy lever for CJK text (source: OpenAI cookbook, validated across 300+ generations).
Output format
Deliver in this order (prompt first — that's what the user copies):
- 📋 The Prompt — the complete 5-block prompt in a code block, copy-paste ready
- Design notes (brief) — recommended type, decomposition summary, model hint, variants
See output-format.md for the exact template.
Quick validation rules
Before delivering, verify:
- Prompt follows 5-block order (CANVAS first)
- Every CJK label wrapped in
THE TEXT READS: "..."
- Total on-image Chinese text ≤ 60 characters
- Font hierarchy specified (≥ 3 tiers)
- Negative constraints included (no garbled text, no Lorem ipsum, no watermarks)
- If target is Midjourney: no Chinese text on image
Full checklist: validation.md. Pitfall fixes: troubleshooting.md.
References (one level deep — read when needed)