← All playbooks
SafetyWe run this
Human approval before anything goes public
Agents draft, sort and flag. A person approves the exact text before it's sent or posted.
Why we run it
Every agent we run that could speak in public follows this rule. It's the cheapest insurance you can buy.
Steps
- List every channel where an agent could say something in public: X, email, DMs, reviews, ads.
- For each one, make 'draft' the agent's last step. Sending is a separate human action.
- Put placeholders where real data is missing: [METRIC], [CONFIRM], [SITE]. Never let the agent fill them with a guess.
- Flag drafts that name a customer, a price or a number, so they get a second look.
- Write the rule into the agent's standing instructions, not just one prompt.
The promptIn the kit
Standing rule for all public output: - You draft. You never publish, post, send or reply on your own. - Where a real number, link or fact is missing, write [METRIC] or [CONFIRM] instead of guessing.
Copy-paste kit · $29Full prompt and config, setup checklist, standing guardrails, updates.
Also in The Revenue Desk, The Full Bench.
Guardrails
- No fake reviews, testimonials, accounts or numbers. Ever.
- Naming a customer in public needs that customer's OK.
What broke
- A launch draft named a client before we had their permission. The flag rule caught it, we asked, and only then kept the name.
Our stack
Written into Hermes and Grok standing instructions.
Want it running on your accounts?We install and adapt it for you, from $500.
Have us install it