Most people arrive at an AI infographic generator holding a report and hoping the tool will find the story inside it. That is the one part it cannot do. What it can do is the part that used to take an afternoon: choose a hierarchy, place the numbers, draw the elements and set the type so the page reads top to bottom.

That split — structure is yours, production is the generator’s — is the whole model. Once you know which half you own, the tool stops feeling like a slot machine.

Three decisions happen before anything is drawn

A good infographic generator is not one model making one guess. It resolves three separate questions, in this order.

1. Layout — the topology of the information. Is this a sequence, a comparison, a hierarchy, a cycle, a map? Our infographic Skill works from 21 layouts, and they are descriptions of information structure rather than visual themes: linear-progression for anything with a then, binary-comparison for a two-sided claim, comparison-matrix when there are several factors, hierarchical-layers for priority, hub-spoke for one idea with satellites, dashboard for metrics, dense-modules when the point is density itself.

2. Style — the visual language. Print register, colour discipline, texture, line quality. The style set we work from is drawn from editorial references rather than generic “modern” prompts, because a style is what decides whether the result reads as a magazine plate or as a stock illustration. Style and layout combine freely: binary-comparison with a risograph look is a different sentence from binary-comparison with a blueprint look.

3. The subject — what the largest element on the plate actually is. This is the one nobody asks about, and it is where most generated pages go flat. A number, a face, a place, a labelled specimen: each demands a different entry point. A biography whose biggest element is an object, not a face, has lost the plot before the palette is chosen. So the subject is not a prompt input — it is read off the content and it constrains the other two.

Two of those you choose. One is derived. A prompt that leaves the third unstated produces the fastest-looking failure in this field: a uniform grid where every tile is equally loud.

The four inputs that produce a usable plate

The generator is only as good as the material you hand it. In practice four inputs separate a plate you edit from a plate you rebuild:

  • One claim, stated in a sentence. “Our support team answers 80% of tickets in under an hour; the industry average is two days.” Every element on the page then has a job: support the sentence or get cut.
  • The numbers with their units and dates. Six hours a week is not the same claim as six hours a month, and a figure without a period is a rumour.
  • The format. A 1080 × 1080 square, a 1080 × 1350 vertical and an A2 poster want different densities from identical data. Say which, because it changes the layout more than the style does.
  • The audience and register. An operations manager and a design lead will accept different amounts of jargon and different amounts of whitespace.

Everything else — the icon set, the colour ramp, the paragraph that introduces the page, the spacing — is production, and production is now cheap.

What the generator gets wrong, and what fixes it

Three failures show up again and again. They are predictable enough that the pipeline is built around them.

Everything is equally important. Left alone, a model weights a supporting statistic like a headline. The fix is to rank the content before the prompt is written: exactly one primary element, carrying roughly 55–65% of the plate’s visual weight, with a type scale that steps at least 20% between adjacent levels. If two sections tie for the top, you have two infographics.

Numbers get invented. Ask for a chart and a model will happily round, extrapolate or dream a plausible date. The fix is to declare, per element, whether it is verbatim (copied character for character), inferred (invented to fit the subject) or derived (reduced from something you supplied). Anything that must survive the render goes in as literal text, and a render that drops it is a failed render, not a near miss.

Text comes back garbled. This is the honest limit of bitmap generation. Short labels and headline figures usually hold; a paragraph set at 8pt does not. Keep on-image text to headlines and labels, and put the explanation in the post copy around the image.

If the numbers must be exact, the copyable version — a table or a plain HTML page — is the better artifact and the image is the advertisement for it.

The pipeline, in the order it runs

It helps to see the five steps a generator actually walks, because it explains which of your inputs matter and when:

  1. Read the material and name the subject. Not the topic — the subject: is the biggest thing on the page a number, a face, a place, or an object?
  2. Rank the content. One primary element, a second tier at 20–30% of the visual weight, a quieter third tier for evidence.
  3. Pick the layout and the style from fixed catalogues, with the aspect ratio fixed before either is chosen. Aspect constrains layout more than style does, so choosing it late forces a redesign.
  4. Assemble the prompt. Every element is tagged: literal text copied verbatim, values inferred to fit the subject, or figures derived from what you supplied. Only the literal ones must survive the render unchanged.
  5. Confirm, then render. A render is the expensive step, and a plate built on the wrong subject wastes it. That is why a good pipeline stops for confirmation before generating rather than after.

Read in order, those steps explain a specific kind of failure: a plate that looks competent and argues nothing. It usually means step one was skipped and the layout decided the argument instead of the content.

Reviewing the render in sixty seconds

Six checks, in this order, catch almost everything:

  • Squint. One element should still dominate. If the page turns into an even texture, the hierarchy failed.
  • Shrink it to about 120 pixels wide. One shape, one direction. Even grey noise means it will die in a feed.
  • Read every number character by character. Check units and dates too. A figure that got rounded is a failed render, not a near miss.
  • Count the accents. One colour marking one thing. Two accents mean two entry points, which means none.
  • Read the second tier aloud in order. If the order is not obvious from the labels, the path through the page is missing.
  • Check the edges. Captions and controls sit over the bottom and right of most social canvases; anything inside 8–10% of the edge is at risk.

If three checks pass and two fail, edit the brief rather than rewording the prompt. The prompt is a translation of the brief; a translation cannot fix a bad source sentence.

A brief that works

This is the shape we ask people to send. It runs to about sixty words and produces something to edit rather than something to rebuild:

Make a one-page infographic for a LinkedIn post. The claim is that remote teams lose about six hours a week to scheduling. Use these three figures exactly: six hours lost, 41% of meetings booked across time zones, three tools in a typical stack. Audience is operations managers. Vertical format, dark text on light ground.

Note what it does not say. It does not name a style, a layout or a mood. Those are the generator’s job, and a brief that names a colour scheme instead of a claim has confused the two halves of the work.

When a bitmap is the wrong output

Three cases where you should stop and build something else:

  1. The content is a table. Nine columns of values belong in a table, not in a plate with a legend.
  2. The numbers will change. A generated PNG is a snapshot. If the figure updates monthly, put it on a page you can edit.
  3. The reader needs to copy or search the values. Text inside an image is invisible to search engines, screen readers and anyone’s clipboard.

For everything else — a launch figure, an explainer, a social post, a poster — the production work is genuinely off your desk. The thinking is not.

How this post was made

Written by Alex Zhang, founder of NeoSpark, from the pipeline the infographic Skill runs: content analysis, a ranked structure with one primary element, a layout and style chosen from a fixed catalogue, a prompt with literal text declared as literal, then a render. The failure modes above are the ones we hit while building it, not a literature review. Where a claim is a judgment — like the 55–65% primary share — it is our composition rule, not a measurement of your audience.

Two related pieces go deeper: how to choose a format for a feed and how to structure a poster people finish. To see the Skill the pipeline came from, start at the infographic designer listing, or browse the rest of the catalogue.