There is no separate style for AI-friendly docs. The same things that make a reference clear for a careful reader make it usable for a model. The difference is that models punish ambiguity more visibly. Where a human guesses and moves on, a model guesses and submits.
A few patterns that help both.
Put constraints up front
For a parameter, the first sentence should contain the type, the unit if any, and the range or set of allowed values. Examples and edge cases come after.
Worse:
expiry: how long the token is valid. Tokens expire after this period.
Better:
expiry: token lifetime in seconds. Integer between 60 and 86400. Defaults to 3600.
A human skimming gets the answer in the first line. A model parsing the page does too.
Use exact identifiers, not paraphrases
If your parameter is user_id, refer to it as user_id in prose. Not "the user's ID," not "the user identifier." A model that sees a paraphrase may use the paraphrase as the field name.
Same goes for status codes, error names, and resource types. Pick the canonical form and reuse it.
Be explicit about units
Seconds or milliseconds. Bytes or kilobytes. UTC or local time. Twenty-four hour or twelve. A human can usually guess from context. A model often cannot. State it once, in the parameter description.
Don't put critical info only in tables
Some models drop or compress tables when context is tight. If a required field, a default value, or a constraint exists only in a table, the model may miss it. Repeat it in prose at least once.
Avoid forward references for required information
"See the authentication section." A human follows that link. A model often does not, especially if the agent has retrieved only the current page.
Anything required to make a call should appear on the page that documents the call, even if it duplicates a paragraph from elsewhere.
A summary
These are not new rules. They are the same things a careful tech writer has always done. Writing for models gives you a strict reader who will not let you skip them.
