GPT Image 2 Knowledge Graph Prompt Guide: 5 Production Templates for Exam Prep, Xiaohongshu, Lecture Notes, Slides & SOPs

Before GPT Image 2, "use AI to make a knowledge graph image" was a meme more than a workflow. Every other model would either spell the headings wrong, jam ten sections into three, or output a pretty-but-empty poster that read like decorative noise. So most people gave up and went back to PowerPoint, Figma, or hand-drawn iPad notes.

That changed this spring. With GPT Image 2, you can now hand the model a single structured prompt and get back a study-grade knowledge card — correct typography, clean section grid, the right arrows in the right places — in one shot. This guide is the prompt framework I keep reusing, plus five copy-paste templates for the five scenarios where readers actually search for this stuff: civil-service exam prep, Xiaohongshu / social knowledge cards, classroom handouts, slide visuals, and operational SOPs.

1. First, what does "AI knowledge graph image" even mean?

Two very different things share that phrase:

• Technical knowledge graph (KG): an entity–relation–entity triple store, the kind you build with Neo4j / RDF / GraphRAG. That's a data structure, not an image.
• Visual knowledge-graph image: a printable / shareable poster that visually organises a topic — definition, types, workflow, pitfalls, examples, mnemonic. This is what people actually want when they search "how to make a knowledge graph with AI."

This guide is about the second one. The first is an LLM + graph-database job and has nothing to do with image models — it's out of scope here.

2. The 5-block universal prompt template

Every successful knowledge-graph image I've shipped with GPT Image 2 follows the same five blocks, in this order. Skip any one and the output collapses into either pretty wallpaper or alphabet soup.

[1] CANVAS — orientation (vertical 1024×1536 or horizontal 1536×1024),
    paper-style background, dominant accent color.
[2] TOPIC HEADER — exact title text + one-line positioning sentence.
[3] SECTIONS — 5 to 8 named cards. Common skeleton:
    Definition → Question Types → Solving Path → Common Pitfalls →
    Comparisons → Worked Example → Mnemonic.
[4] VISUAL — line weight, icon style, arrow style, palette accents,
    rounded-card vs. ruled-notebook aesthetic.
[5] CONSTRAINTS — "THE TEXT READS: ..." quoting for every label,
    no extra logos, no decorative filler, preserve exact spelling.

Why these five and not six or four:

• Canvas first. The model commits to layout before content. If you describe the topic before the canvas, you get random aspect ratios.
• Header before sections. The H1 anchors hierarchy. Without it, all section cards render at the same visual weight and the image reads flat.
• 5–8 sections is the sweet spot. Below five looks empty; above eight starts losing labels even on GPT Image 2.
• Visual block last among the descriptive ones. Style descriptions placed early get over-applied; placed late they act as a finishing pass.
• Constraints at the bottom, in imperative voice — the model treats them as guardrails rather than suggestions.

One trick worth repeating from the OpenAI cookbook: every piece of on-image copy should be wrapped as THE TEXT READS: "...". That single phrase is the biggest accuracy lever for non-Latin scripts and long titles.

3. Scenario 1 — Civil-service exam study card (the headline example)

This is the one I'd start with, because exam-prep knowledge cards are the strictest test: they need accurate Chinese, dense information, a clear solving workflow, and a "study-handout" aesthetic that doesn't look AI-generated.

Here's the prompt that produced the cover image of this post (a study card for "Civil-Service Interview Analysis Question Framework"):

Create a vertical (1024×1536) civil-service exam study infographic
on the topic "面试综合分析题答题结构 / Civil-Service Interview
Analysis Question Framework". The goal is to help a reader
understand: what this question type tests, how to recognise it,
the standard answering workflow, common pitfalls, similar
question types, a worked example, and a memorable mnemonic.

CANVAS: clean light paper background, deep-navy title block,
charcoal body lines, with restrained accents in blue, teal, gold,
and a single red flag for warnings.

LAYOUT BLOCKS (rounded cards with thin borders, numbered tags):
1) Header: exact title + one-line positioning.
2) Core Definition — what this question actually tests.
3) Question-Type Signals — prompt phrasings and keyword tells.
4) Solving Path — 4 to 5 numbered steps connected by hand-drawn arrows.
5) Common Pitfalls — trap options, wrong reasoning patterns.
6) Comparison — distinguish from adjacent question types.
7) Worked Example — short stimulus + correct decomposition.
8) Mnemonic — one short rhyming line or three-keyword summary.

VISUAL: looks like a high-quality printed exam handout crossed
with a hand-illustrated education poster. Use rounded cards,
thin rule lines, numbered tags, hand-drawn arrows, small
zoom-in callouts, and a dedicated "易错提醒" warning strip.

CONSTRAINTS: every visible heading and body string is rendered
verbatim — THE TEXT READS exactly what is specified. No extra
logos, no Lorem ipsum, no decorative filler text. Information
density is high but the page stays balanced and uncluttered.

A few things worth noting:

• I write the Chinese title plus an English gloss. GPT Image 2 handles either, but bilingual prompts make the model treat the topic as serious / educational rather than meme-y.
• Numbered tags + hand-drawn arrows make the output read like a study handout, not a marketing flyer.
• The "easy-mistake / pitfalls" card is the single most underused element in AI knowledge cards. Adding it is what makes the image feel authored rather than generated.

You can browse the rendered output and 20+ similar templates on the /prompts page — every card there is reproducible from its source prompt.

4. Scenario 2 — Xiaohongshu / social-media knowledge card

For Chinese social platforms (especially Xiaohongshu / RED), the de-facto spec is 1080×1440 vertical, F-shaped reading flow, three-card series. The prompt below produces a single card from that series; reuse it three times with the same accent palette to get a swipe-through set.

Create a vertical 1080×1440 knowledge card for Xiaohongshu titled
"3 Habits That Actually Lower Cortisol".

CANVAS: warm cream background (#FFF8EE), subtle paper grain.
A single accent color of muted terracotta (#C96E5A).

LAYOUT:
- Top: progress label "认知觉醒 · 卡片 2 / 3" in tiny mono caps.
- Headline block: H1 in handwritten serif, sub-line in clean sans.
- Middle: three numbered "habit" rows. Each row has a small flat
  icon on the left, a 3-word habit name, and a 1-sentence "why".
- Bottom: a single quote strip with a hand-drawn underline.
- Right edge: tiny vertical text "swipe →".

VISUAL: minimal flat illustration, two-color palette only, 1pt
hairlines, slight paper texture, generous whitespace, F-shaped
reading flow (title top-left, scan down).

CONSTRAINTS: text rendering must be pixel-perfect — THE TEXT
READS each label exactly as written. No watermark. No platform
logos. Output looks like a published Xiaohongshu carousel card,
not a Canva template.

Two retention tricks specific to Xiaohongshu:

• The "X / Y" progress label tells viewers there are more cards — it raises swipe-through and average view duration, both of which the algorithm rewards.
• A handwritten H1 paired with clean sans body is the visual signature of "thoughtful creator" rather than "agency content," which is the entire pitch of the platform.

5. Scenario 3 — Classroom handout / lecture note

For teachers and tutors. The trick is that a handout is wider than a knowledge card — it goes horizontal, has a left navigation column, and lets the body breathe.

Create a horizontal 1536×1024 classroom handout titled
"Photosynthesis — One-Page Lecture Note for 8th-grade Biology".

CANVAS: graph-paper background, navy header bar, hand-drawn
margin notes in pencil-grey.

LAYOUT:
- Left rail (20% width): table of contents with 5 numbered items.
- Main area (80% width) split into three rows:
  Row A — annotated diagram of a leaf cross-section with arrows
    pointing to chloroplast, stomata, xylem, phloem.
  Row B — the balanced chemical equation in a boxed callout,
    with sun, water, CO2, glucose, O2 labeled as small icons.
  Row C — a 4-step "Light vs Dark Reactions" comparison table.
- Bottom-right: a "Common Misconceptions" red flag box with
  three bullets, each one sentence.

VISUAL: pen-and-marker textbook illustration style. Pencil-grey
linework, navy headings, accent yellow for highlights, red for
warnings only. Friendly but not childish.

CONSTRAINTS: every label is rendered verbatim. The handout reads
like it was prepared by a science teacher, not designed by a
marketing agency. No school logos, no watermarks.

If you're producing a whole curriculum, generate one handout, then in follow-up turns ask GPT Image 2 to preserve the layout and only swap the content block — its local-edit fidelity is what makes batch production economically viable now.

6. Scenario 4 — Slide-deck visual (boardroom-grade)

The classic painful slide elements — TAM/SAM/SOM concentric circles, 2×2 matrices, swim-lane diagrams — used to take 30 minutes each in Figma. They take one prompt now.

Create a horizontal 16:9 slide visual titled
"2026 GTM Plan — Where We Win This Quarter".

CANVAS: pure white background, single brand-blue accent (#2563EB),
charcoal text. Boardroom presentation aesthetic.

LAYOUT: 2x2 matrix.
- X axis label: "Effort (low ← → high)"
- Y axis label: "Strategic value (low ← → high)"
- Quadrants, each with a 2-word verdict tag and 2 example bullets:
  Top-right "DO NOW": Enterprise pilot, Channel partner.
  Top-left "QUICK WINS": Lifecycle email, Pricing test.
  Bottom-right "RECONSIDER": Conference circuit, New SDR hire.
  Bottom-left "PARK": Affinity merch, Side-project blog.
- Footer micro-text: "Source: Q2 OKR review, internal".

VISUAL: clean sans-serif, 1pt rule lines, generous whitespace,
no decorative icons. Reads as a McKinsey-style consulting slide.

CONSTRAINTS: every quadrant label is rendered verbatim. No
watermarks. No fake company logos. The image looks like a
PowerPoint screenshot, not a stock illustration.

Pair this prompt with a follow-up like "Now generate a TAM / SAM / SOM concentric-circle visual in the same visual language" and you have a coherent deck instead of a Frankenstein of screenshots.

7. Scenario 5 — SOP / workflow diagram

Operations teams have the cleanest ROI on this entire workflow — one good SOP diagram saves an hour of Visio time and reads better than the manual it replaces.

Create a vertical 1024×1536 SOP diagram titled
"Customer Refund Process — Standard Workflow (2026)".

CANVAS: white background, navy header strip with the SOP title
and a small "Doc ID: OPS-014 · v2.1" badge.

LAYOUT: top-to-bottom flowchart with 6 numbered nodes connected
by directional arrows.
1) "Customer submits refund request" — owner: Support L1.
2) "Verify order + payment status" — owner: Support L1.
3) Decision diamond: "Within 30-day window?" — Yes → step 4,
   No → step 6.
4) "Approve refund in Stripe dashboard" — owner: Finance.
5) "Send confirmation email + close ticket" — owner: Support L1.
6) "Escalate to Support L2 for exception review" — owner: L2.
Right side: a sidebar with three "Common pitfalls" warnings
and the SLA: "Refunds processed within 24h business time".

VISUAL: clean Visio-style flowchart, rectangles for actions,
diamonds for decisions, hand-drawn-feel arrows in navy, owner
tags in small pill labels. No icons, no illustration.

CONSTRAINTS: every node label and owner tag rendered verbatim.
No watermarks. Output reads like a print-ready company SOP page.

A subtle pro move: ask the model to render the document ID and version number in a small badge. Reviewers treat documents with version metadata as authoritative; it's the cheapest credibility upgrade in this entire workflow.

8. Six common pitfalls (and how to fix them)

After ~300 knowledge-card generations, these are the failure patterns I now pre-empt in every prompt:

If you're getting one of these and it's not on this list, open the prompt library and reverse-engineer the closest working example — that's usually faster than debugging from scratch.

9. Sizing cheat sheet

For people who only want the dimensions and a one-line prompt suffix:

The aspect ratio alone changes whether the model commits to a "card" or a "page" mental model. Use the column literally — don't translate 1024×1536 into "vertical infographic" and expect the same crispness.

The bottom line

The interesting shift isn't that GPT Image 2 "looks better." It's that one-shot, ready-to-publish knowledge images are now a workflow rather than a fantasy. The five scenarios above cover the bulk of what people actually search for when they Google "how to make a knowledge graph with AI" — and the same 5-block prompt template carries across all of them.

If you want to skip writing prompts and just see what GPT Image 2 can do, the /prompts page has 20+ production-ready templates (including the civil-service card from this post) you can fork, and /explore has the full output gallery with source prompts attached.

Further reading

• Did GPT Image 2 Really Dethrone Nano Banana? My Verdict
• GPT Image 2 vs Nano Banana 2 vs Midjourney v7 — A Production-Focused Comparison
• GPT Image 2 Prompt Writing Guide — 7 Rules for 90% Hit Rate