Agent guardrails: approved claims, banned phrases and disclosures

Where to find it

1. 1Open Project settings for the project.
2. 2Choose the Guardrails tab.
3. 3Add your rules, choose what happens when one is broken, and press Save guardrails. Agents use them from their next publish.

The kinds of rule

• Approved claims: statements your team has signed off, such as prices, guarantees, certifications and results ("Professional costs €269 per month"). Pick a type for each (pricing, policies, company facts…) and, if it is only true in some countries, list those markets. When an agent checks a draft, approved claims count as the strongest evidence, above your knowledge base.
• Banned phrases: words or claims that must never appear, such as "guaranteed" or "#1 in Europe". Add the reason (agents see it) and, if you like, the wording to use instead. Phrases match whole words in any language and any capitalisation, also inside headings and formatted text, so "guaranteed" does not match "unguaranteed".
• Required disclosures: text that has to be in the copy, for example "#ad" on sponsored social posts or a regulatory notice in one country. Choose the channels (articles, landing pages, social posts, ads, email), markets and languages it applies to; leave them empty to apply it everywhere. Matching ignores capitals and extra spaces.
• Brand rules: plain-language guidance every agent follows when it writes (“Call it a platform, never a tool”). They are never checked or flagged. [More on brand rules](/dashboard/help?article=brand-rules)

Warn or block

• Off: agents publish without checks.
• Warn (the default): agents publish, and are told what broke a rule so they can fix it and tell you.
• Block: an agent cannot publish while a blocking rule is broken. It gets the list of problems, fixes them and tries again. The attempt appears in Agent activity as *Blocked by guardrails*.

• Ads: when the assistant creates a Google responsive search ad, or switches a Google or Meta campaign on, the ad text (headlines, descriptions, body, call to action of the campaign’s creatives) is checked first. Pausing a campaign is never blocked.
• When you publish yourself: Broadcast Review shows guardrail warnings under each caption, and Content Studio shows them above the publish options. They update as you type, and Check claims too runs the claim check. You are never blocked; the warnings are there so you can fix things before they go out.

The claim check

• Approved claim: matches one of your approved claims.
• Supported: your knowledge base backs it.
• Contradicted: your approved claims or knowledge base say something different, for example "£29 a month" when the approved price is £49. The check shows what the evidence actually says. This blocks publishing when enforcement is Block.
• No source: nothing you have addresses it. For the claim types you mark as sensitive (pricing and policies by default) this is a warning: remove it, reword it, or add it to your approved claims once you have checked it.

Test a draft yourself

1. 1In Project settings → Guardrails, scroll to Test a draft.
2. 2Paste a post, article or ad, and choose the channel (and a market, if it is not the project’s country).
3. 3Leave Check factual claims with AI on to include the claim check, then press Check draft.
4. 4You see what an agent would be told: each broken rule with the passage and a suggested fix, and each claim with its result. The test uses the saved rules, so save changes first.

Change guardrails by asking an agent

• Only owners and admins can change guardrails, whether in settings or through an agent.
• What agents may change is set in the Guardrails tab: All changes (the default), Stricter only, or Off. With *Stricter only*, agents can add banned phrases and disclosures and raise enforcement, but anything that loosens the rules (removing a rule, lowering enforcement, adding an approved claim) has to be done in settings.
• Adding an approved claim counts as loosening, because it changes what the claim check accepts as true.
• Every change is recorded in Agent activity, with what changed and whether it loosened the rules. A change the setting refused is recorded as *Blocked*.

For agents and MCP clients

• get_project_guardrails: read the project’s rules before writing, with each rule’s id and what agents may change.
• update_project_guardrails: add or remove rules (by id or exact text), change enforcement or sensitive claim types; dryRun previews without saving. Refused with agent_edit_not_allowed when the project’s setting does not allow the change.
• check_draft_guardrails: check a draft (text, channel, optional market and language; the AI claim check is on unless claims is false). "Would be blocked" means publishing will be refused.
• A publish refused by the rules returns the error guardrail_blocked with the list of findings.
• The Reaudit desktop agent and assistant already follow these steps. With Claude or ChatGPT, ask them to "check this against our guardrails before publishing".