Most teams that ground an AI agent start with the style guide, and that's the right first move. It teaches the agent your voice, your terminology, whether you write "log in" or "sign in," how you capitalize headings. We've written before about grounding an agent in your style guide, and it makes drafts feel like they came from your team.
But a style guide governs words. It says almost nothing about structure. And structure is where a draft either slots cleanly into your docs or reads like it wandered in from another site.
Voice and shape are different problems
Picture two drafts of an authentication page. Both use your exact terminology and tone. One opens with a prose paragraph, then a code sample, then a list of error codes. The other opens with the error codes, buries the example halfway down, and never explains where to get a key.
Both passed the style guide. Only one matches how your authentication pages actually work. The style guide had no opinion on the order, because order isn't a voice problem. It's a shape problem, and shape is what your readers have learned to expect from your docs.
An agent that only knows your voice will get the words right and the shape wrong, and the shape is what readers notice first.
What a page template encodes
Your docs already have templates, even if nobody wrote them down. Your endpoint pages probably all follow the same skeleton. Your guides have a recognizable rhythm. Making that implicit structure explicit is what gives an agent something to match.
A useful template captures things like:
| Element | What the agent needs to know |
|---|---|
| Section order | What comes first, what comes last, what's required |
| Required pieces | Every endpoint page has a request example and an error table |
| Depth | How much explanation before the first code block |
| Cross-links | Auth pages always link to the key-management guide |
None of that is voice. All of it is the difference between a draft that drops into place and one you have to rebuild.
Show the agent a good page, not just rules
The fastest way to teach structure isn't a spec. It's an example. A single well-formed page of each type tells an agent more about your conventions than a page of written rules, because it shows the order, the depth, and the linking all at once, in context.
This is the same instinct behind using existing docs as context: your best pages already encode your conventions. Point the agent at them as templates and it has a concrete target to match, not an abstract checklist to satisfy. When you add a new page type, add a canonical example of it, and the agent picks up the pattern without anyone writing a new rule.
Where templates and voice meet
You want both. Voice without structure gives you on-brand drafts in the wrong shape. Structure without voice gives you correctly shaped drafts that don't sound like you. Together they produce a draft that a reader can't tell was machine-generated, because it lands in the right place and reads in the right tone.
The good news is they're additive, not competing. The style guide handles the sentence. The template handles the skeleton. An agent grounded in both has very little room left to improvise.
How ReadMe puts this together
ReadMe's AI agent works from your existing docs, which means your real pages double as the templates. When your endpoint pages share a shape, the agent's drafts inherit that shape. Pair that with style-guide grounding and the drafts match your docs on both axes, voice and structure, before a human ever opens the branch.
The takeaway: don't just tell an agent how to write. Show it what a finished page of yours looks like, and let the shape do the teaching. Start a project to ground an agent in the docs you already have, or talk to us about Enterprise for team-wide controls.
