How we write and test

What we publish

Three kinds of pages carry editorial content. Guides explain how to get something done with the tool, such as picking an export size or sharing a prompt safely. Prompt examples each take one complete prompt and explain why it is built the way it is. Template and platform pages describe one template, or the kinds of prompts people write for one AI tool, and which card suits them.

Written from the tool itself

Anything we say about how an image looks is checked against what the export actually draws: the colours, the fonts, the canvas sizes, where the text sits, and when a long prompt gets shortened. When the description and the export disagree, the description is wrong and we change it.

Images labelled as exports are real exports. To make one, a scripted browser opens the public editor, pastes the prompt, leaves the template’s defaults alone and clicks Download PNG, exactly as a visitor would. The file is recompressed without changing a single pixel and published unedited. The smaller cards drawn on some pages to compare templates side by side are simplified mock-ups, not exports; each template’s own page shows a real exported PNG.

A worked example: the export size matrix

The template and export size matrix is the most detailed test we have published. On September 12, 2026 we exported three test prompts (short, long and structured, and code-heavy) through all eight templates at all four sizes: 96 exports, and all 96 produced a valid PNG with no text shortened. A supplementary run with a deliberately oversized prompt of 3,366 characters made 32 more exports; the editor shortened the text in 8 of them, all at 1200 × 675, and fitted it everywhere else.

That test also found a real bug. The live preview cut long prompts off even when the exported PNG did not. The preview has since been rebuilt to use the same renderer as the export, and the guide now carries a dated note saying which of its tables describe the old behaviour. Publishing that note, rather than quietly replacing the table, is how we handle results that stop being true.

How prompt examples are written

We write every example prompt for this site, based on everyday tasks: turning meeting notes into actions, reviewing a pull request, briefing a coding agent, revising for an exam, rewriting in a house style. None of them is collected from visitors, and no published prompt from another site is copied or reworded. Each page explains the job of every instruction, so you can tell which lines matter before you edit it.

What we do not claim

• We do not publish AI model answers as evidence that a prompt works. Answers vary by model, version and run, and a cherry-picked response proves little. The examples explain the reasoning instead.
• We do not rank or review AI products. Platform pages describe the kinds of prompts people write for a tool and how they look as a card, not how good the tool is.
• AI PromptShot is independent and is not affiliated with OpenAI, Anthropic, Google, Cursor or Midjourney. Nothing on this site is official guidance from them.
• We do not take payment to mention, link to or feature anything. The site is funded by display advertising only; see the editorial policy.

Checks that run before anything ships

Some rules are enforced by automated tests in our codebase, so a page cannot be published while it breaks them. The template and example pages must stay within a set length, must not share sentences with each other, and must avoid stock filler phrases. Platform pages must not repeat each other’s paragraphs. A separate test blocks wording that would misdescribe how your prompt is stored, such as claiming nothing is ever saved when the Save button exists.

When the tool changes

When the editor or the export changes, we update the pages that describe it, and the “Last updated” date on each changed page moves. What counts as an update, and how corrections are recorded, is set out in the editorial policy. If you spot something on the site that no longer matches the tool, tell us.