For Agents

The idea

An agent that writes HTML, generates an image, or drafts a paragraph is a maker, and the same rule applies: it is shipping to someone it will never meet. Most accessibility guidance assumes a human reader with time to interpret it. The guidance here is written so that an agent can act on it directly, and so that a person can check what the agent did.

What will be here

• Human-readable guidance. How to decide on alt text, how to check a focus state that static analysis cannot see, what to verify before returning a component.
• Machine-readable versions. The same rules at /llms.txt, /agents.md, and as a JSON ruleset with ids, severity, and a test method for each rule.
• A test corpus. Small pages with planted, labeled violations and an expected-findings file, so an agent's accessibility checks can be scored for precision and recall.
• Prompt patterns. Phrasings that reliably produce accessible output, and the ones that quietly do not.

Three rules to start with

1. Alt text is a decision about context, not a description of pixels. Ask what the image is doing on this page before describing what it contains. If it does nothing, the alt is empty.
2. A focus state must be visible on every background it can land on. One outline color is not enough. Use two tones or an offset gap, and never remove the outline without replacing it.
3. Verify, do not assume. Before returning HTML or CSS, check contrast for every text and non-text pair you introduced, and tab through what you built.