Evidence-Locked Infographic Prompt

This is a Prompt for turning an approved evidence packet into a short visual explanation. It makes the audience, lesson, facts, modules, relationship, literal text, and exclusions explicit before image generation. The goal is not to make an infographic look authoritative; it is to prevent the design from saying more than the source supports.

What this Prompt changes

A weak infographic request mixes content and decoration:

Make a professional infographic about checking claims.
Add useful statistics, icons, and a modern layout.

The model must decide which claims matter, invent or retrieve “useful” statistics, choose a relationship, write labels, and design the page. Even when the result looks polished, the reader cannot tell which parts came from the brief and which parts were improvised.

The Evidence-Locked Infographic Prompt separates those decisions. The operator supplies the facts and exact words. The model chooses a visual composition inside declared boundaries. A human then compares the image with the same packet.

Build the evidence packet first

Do not begin with colors. Begin with what the image is allowed to teach. A compact packet is usually enough:

AUDIENCE: [who will read it and what they already know]
LESSON GOAL: [one sentence]

VERIFIED FACTS
- [fact or claim] — source ID: [SOURCE-ID]
- [fact or claim] — source ID: [SOURCE-ID]

TITLE: [short literal title]

MODULES
1. [label] — [meaning] — [reviewable visual cue]
2. [label] — [meaning] — [reviewable visual cue]
3. [label] — [meaning] — [reviewable visual cue]

RELATIONSHIP: [process / comparison / hierarchy / cycle / cause and effect]
TEXT ALLOWLIST: "[exact string]", "[exact string]"
EXCLUSIONS: [misleading symbols, adjacent concepts, unsupported details]

A source ID belongs in the review record even when it does not appear in the art. That lets an editor trace a module back to evidence without forcing a long URL or footnote into generated typography.

Give every module one lesson job

Three to six modules are usually enough for a single image. Each module needs a short label, a meaning, and one visible cue. “CHECK — determine whether the source supports the claim — magnifier over a link” is reviewable. “CHECK — make it insightful” is not.

Then name the relationship. A process needs a sequence and connectors. A comparison needs aligned categories and balanced visual weight. A hierarchy needs a clear parent and child structure. A cycle needs a return path. If no relationship is declared, a generator can satisfy every module while failing to explain how they fit together.

Keep generated wording short and literal

OpenAI's image-prompting guidance recommends making requested text explicit and structuring detailed prompts clearly. This Prompt uses a literal allowlist: every permitted string is quoted, requested once, and everything else is banned.

Keep the list short. Generated text is appropriate for a title, one- or two-word module labels, and brief captions that are checked manually. It is a poor place for legal qualifiers, citations, long definitions, or dense numeric tables. For those, ask the model to leave a blank panel and add editable type in a design tool.

TEXT
Leave every text area blank. Render no letters, numbers, symbols,
signatures, captions, pseudo-writing, or watermark. The approved title,
labels, values, and citations will be added as editable typography later.

Our filled test fixture

The test topic was deliberately synthetic and non-statistical. It explains a four-step editorial gate rather than asserting a claim about a company, product, health outcome, or market. The ten allowed strings were:

"FROM CLAIM TO DECISION"
"SOURCE"
"Owner + date"
"CLAIM"
"Quote the source"
"CHECK"
"Support found?"
"DECISION"
"Cite or hold"
"HOLD IF UNSURE"

The prompt asked for a vertical 4:5 poster, four equal top-to-bottom modules, one connecting arrow path, warm off-white paper, deep navy structure, muted coral accents, mustard highlights, and no other words, numbers, logos, or watermarks. The exact filled request is preserved in the public JSON record.

What two identical calls produced

The variation is not a defect by itself. Both images teach the same sequence. It does mean that a brand system cannot rely on one prompt to preserve exact component geometry, icon styling, color placement, or line breaks.

A practical review gate

Review content before style. Use the same packet that produced the image:

1. Fact gate: every visible claim must map to a supplied fact. An attractive extra statistic is a rejection.
2. Module gate: every requested module appears exactly once and has the intended meaning.
3. Relationship gate: the viewer can see the sequence, comparison, hierarchy, cycle, or causal structure without reading a long explanation.
4. Text gate: compare every visible character with the allowlist. Reject misspellings, duplicates, substitutions, and pseudo-writing.
5. Exclusion gate: inspect for logos, misleading icons, unwanted adjacent concepts, and fabricated citations.
6. Legibility gate: check the actual delivery size, not only a zoomed original.
7. Publication gate: add citations, accessibility text, editable typography, and specialist review outside the image when the use case requires them.

Repair only the failed invariant

When most of the image is correct, avoid rewriting every instruction. Name the failed invariant and preserve the rest:

Edit only the failed text layer.
Keep the canvas size, module order, connectors, icons, palette,
spacing, and all correctly rendered words unchanged.

Replace the incorrect phrase with exactly: "[CORRECT PHRASE]"
Remove: "[UNWANTED OR DUPLICATED PHRASE]"
Add no other letters or numbers.

Use an editing route that supports image inputs when available. OpenAI documents both generation and editing for GPT Image 2, while its broader image guide distinguishes single-image Image API work from conversational multi-turn editing in the Responses API. Record which route was actually used instead of assuming that a chat edit and an API edit behave identically.

When not to use this Prompt

• Do not use it to discover facts or citations. Research and approve the packet first.
• Do not use it for precise quantitative charts. Generate the chart from data with deterministic software, then use illustration only around it.
• Do not put long compliance language inside generated pixels. Add accessible, editable text afterward.
• Do not treat a medical, legal, financial, or safety graphic as approved without qualified domain review.
• Do not claim a model-specific result when the generation tool does not reveal its model.
• Do not infer a success rate from these two retained images.

Why this came from a gallery review

The Awesome GPT Image 2 review found a valuable progression: many concrete examples can become a smaller template library, and templates can become an Agent workflow. Grow AI Skills adds one more layer: an evidence contract and a public test record.

The upstream collection inspired the use-case category, not the wording of this Prompt. We did not republish an upstream image, copy a case, or rename the upstream Skill. That distinction protects provenance and makes the resulting asset easier to maintain.

Final use rule

A successful image call is a candidate, not a publication decision. Freeze the facts, declare the relationship, limit the words, generate, inspect the original file, and keep the result on hold if any claim or label cannot be verified. The best infographic is not the one that looks most certain. It is the one whose visible meaning can be traced back to an approved packet.