Skip to main content
AI agents read the docs through the site’s llms.txt index, per-page Markdown, and the docs MCP server. They often see only part of a page. Follow these guidelines so they get the right answer from whatever part they read. See Docs for AI Agents for the formats agents use.

Write a description for every page

The llms.txt index lists each page by its title and description. Agents use it to decide which pages to read.
  • Add a description to the frontmatter of every page.
  • Write one sentence that says what the reader can do or learn.
  • Start with a verb when the page is a task.
  • Don’t repeat the title.

Keep the answer in the main content

Content inside tabs and accordions can be lost or flattened when a page is converted to text.
  • Put the core answer or required steps in the main flow of the page.
  • Explain the general concept before the tabs, then use tabs for variations, such as the same step in different SDKs.
  • Use accordions only for optional or supplementary detail.

Make each section stand on its own

Search and retrieval tools often return one section, not the whole page.
  • Name the subject in each section. Write “Session Replay,” not “it” or “this feature.”
  • Repeat key conditions where they apply, such as the plan or Data Residency region.
  • Don’t write “as mentioned above” or “see the previous section.” Link to the section instead.
  • Write headings that describe the section’s content.

Be explicit and consistent


Write code that runs as written

Agents copy code samples directly.

Handle deprecated and outdated content

Agents can’t tell an old page from a current one unless the page says so.
  • State that a feature is deprecated at the top of the page, with the replacement and dates.
  • Add noindex: true to the frontmatter of a page that is fully replaced by another page. This removes it from search and AI indexes. The page still works for anyone with the link.
  • Remove or redirect pages for features that no longer exist.

Write prompts readers can give their agent

When an agent can complete a task for the reader, such as installing an SDK, give the reader a prompt to copy.
  • Put the prompt in a txt code block.
  • Keep it short and specific to one task.
  • Include the placeholders the reader needs to replace, such as YOUR_PROJECT_TOKEN.
  • Tell the reader what to check after the agent finishes.