Thiemo Müller, Founder, Content Sync · · 7 min read
Rules Agents Can FollowAI Content Governance
Most companies already have their rules. They live in brand guides, legal guidance, claims libraries and market policies. They were written for experienced people, who fill gaps with judgement and know the context.
An AI agent does not have that judgement, and it only knows what is in front of it. If you want agents to write content that is close to compliant from the first draft, the rules have to be clear enough to follow without the context in a colleague's head.
These six habits help. None of them needs a tool, and they also make rules easier for people to apply and audit.
1. One Rule, One Idea
A sentence like "Our tone is confident but humble, and we never overpromise or use jargon" holds four rules. Split it. "Do not promise outcomes the product cannot deliver" is one rule. "Explain a technical term the first time it is used" is another.
One idea per rule makes it clear what was broken when something fails, and it lets you change one rule without touching the others.
2. Say What a Passing Piece Looks Like
"Avoid misleading claims" tells a writer what to avoid, not what to do. Add the positive form: "Every claim about effectiveness names the study it comes from." A rule that describes a passing piece is easier to follow and easier to check.
3. Scope the Rule
Most rules do not apply everywhere. A rule about promotional statements may apply only to prescription products in the United States. A rule about formal address may apply only to German content.
For every rule, write down where it applies:
- Market: which countries or regions.
- Product: which products or product lines.
- Channel: web pages, email, social posts, print.
An agent that receives only the rules for what it is writing has less to get wrong.
4. Build a Library of Examples
Examples remove more ambiguity than definitions, but one example that passes and one that fails is not enough. Each rule needs a library: about 20 to 50 examples per check, each with the reasoning behind its verdict, reviewed by a person who owns the rule.
- Clear cases: "The healthiest breakfast in America" fails; "A breakfast with 12 g of protein per serving" passes.
- Borderline cases: the hardest ones belong in the library most, because that is where an agent and a reviewer would disagree. "Our most popular breakfast" passes only if the sales figures behind it are on file.
- The reasoning: each example says why it passes or fails, so the next person who maintains the rule knows what it was meant to catch.
Add a new example whenever a reviewer overrules a result. The library grows where the rule was unclear.
5. Name the Evidence
Claims rules are only as good as the evidence behind them. When a rule says a claim needs support, say what counts as support and where it lives: an approved claims library, a study reference, a partner agreement.
If the evidence cannot be found, the claim cannot be made. Writing that down turns a judgement call into a clear step.
6. Give Every Rule an Owner and a Version
Rules change. Regulations are updated, brands evolve, a market adds a requirement. Give each rule an owner who can change it, a version, and the date it took effect.
When someone asks why a page from last spring says what it says, you can point to the rules that were valid when it was published.
Where to Start
Pick one content type in one market, for example product pages in Germany. Collect every rule that applies to it, rewrite them with these six habits, and test them on ten real pages. You will find gaps and duplicates quickly, and you will have a template for the next market.
If you want to see how rules like these apply across every tool your teams and agents use, read how Widebound works. For the bigger picture, start with our point of view.
Drafted with AI agents under our brand and claims rules, checked by Widebound, and read by a person before publishing.