An accessibility audit prompt builder turns an LLM into a fast first-pass reviewer for the content-level parts of WCAG compliance. Most accessibility failures that AI can catch are in the words and structure: missing or useless alt text, headings that skip levels, vague link text like “click here,” and prose written above the reading level of the audience. This tool captures what you are auditing, the WCAG level you target, and your audience’s needs, then builds a prompt that produces a structured, criterion-by-criterion review you can act on.
How it works
You choose the content type (web page, document, email, or app screen text), set the WCAG level — A, AA, or AAA — and describe your audience’s access needs. The builder writes a prompt instructing the model to review the content against the relevant success criteria: alt-text quality, logical heading structure, descriptive link text, reading level, plain-language clarity, and meaningful sequence. It asks for findings grouped by severity with specific fixes, not vague advice. The prompt is generated locally; you paste it alongside your own content into your LLM.
What AI can and cannot check
This is the most important boundary to respect. An LLM can evaluate everything that lives in the text and markup it can see — and it does this quickly and consistently. What it cannot do is test the live, interactive layer: keyboard focus order, screen-reader announcements, color contrast ratios, motion and timing, and form error handling. Treat the AI audit as the content half of the job. Pair it with an automated tool such as axe-core for technical checks and with manual testing using a real screen reader for anything you intend to certify as compliant.
Tips and examples
Set the WCAG level to AA unless you have a specific reason to go higher, since AA is what most accessibility law requires. Be concrete about your audience — “older users with low vision” or “readers whose first language is not English” changes which criteria the audit weights. Ask the model for the fix, not just the flaw; the prompt is written to demand a corrected version of each issue so you can apply changes directly. After acting on the content findings, run the technical and manual layers before claiming conformance, and keep the prompt consistent across pages so your whole site is judged by the same bar.