Shapes of Intelligence / Reference

Bot Style Guide

How to write pages that keep their voice for humans while staying easy for agents to parse, summarize, and act on.

This site has two jobs at once:

  1. A human should be able to read it and feel a point of view.
  2. An agent should be able to read it and recover the structure without guessing.

The mistake is to pick one. The better pattern is: voice in the body, structure at the edges.

The main rule

Do not flatten a distinctive page into SEO paste just to make it machine-readable.

Instead:

That preserves the reason a human would care while giving an agent a stable spine to hold onto.

For overview pages

Overview pages are for orientation, not execution. They should answer:

Use a visible summary block near the top. Prefer plain language over clever labels. If the body is stylized, the summary should be the straight version.

For guide pages

Guide pages tell an agent what to do. They need tighter structure:

If the page contains agent instructions, they should be visible in the page. Hidden comments are acceptable as a mirror, not as the only place the instructions live.

What agents struggle with

These are the patterns that make pages harder to read with rg, extract, or summarize:

Good defaults

A good page skeleton

For a public page on this site, the default shape is:

  1. Title
  2. One-sentence promise
  3. Visible summary or handoff block
  4. Main body
  5. Related links / next steps

That is enough for both a human skim and an agent skim.

Trust model

Telling an agent to read a page is close to giving it code.

That means:

The standard

A page passes if: