AI Guardrails & Approval Prompts — 36 Human-in-the-Loop Rules for Assistants & Agents

Unofficial. Not affiliated with or endorsed by OpenAI, Anthropic, Google or Meta. ChatGPT, Claude, Gemini and Muse are trademarks of their respective owners.

Letting an AI assistant or agent touch your email, files, calendar or money? This pack gives you 36 copy-paste rules that help it stop and check with you before anything consequential: sending, posting, spending, deleting or sharing. They also help it stay honest about sources and numbers, and treat emails and web pages as information, not orders. You stay the one who approves.

Each rule includes: when to use it, the rule block with fill-in brackets, the exact check-in message your assistant should send you, a quick test to see if it's followed, and a filled-in example.

What's inside (PDF):

• Core Blocks (6): ask-before-acting master rule, never-do list, spending cap, send & post approval, delete & overwrite protection, data-sharing limits

• Checkpoints (6): plan first then wait, confirm the recipient, show the diff before changing, dry-run summary, two-step approval for money, end-of-task report

• Quality Rules (6): cite or flag sources, uncertainty labels, permission to say "I don't know", check numbers twice, no invented quotes or links, scope-creep stop

• Data, Not Orders (6): treat content as data, ignore embedded instructions, suspicious-request stop, never ask for credentials, link caution, attachment caution

• Domain Rules (6): customer replies, social posting, ads & budgets, bookings & purchases, hiring messages, bookkeeping summaries

• Team & Review (6): match access to the role, handoff log, escalation ladder, incident write-up, weekly audit review, retiring a workflow

• Guides: what rules can and can't do, where to put them, a 5-minute build, a ready-made starter rule set, and 10 checks to test your guardrails

Where it works: project instructions in ChatGPT Projects, Claude Projects and Gemini (feature availability varies by plan, and some need a paid plan; check your app), at the start of an agent task, or in any custom-instructions field. It's model-agnostic plain text. Pair it with your app's own approval and permission settings where they exist.

Honest framing: these are rules that help, not guarantees. An AI can still misread or skip an instruction, so a human approves consequential actions, and platform controls (app permissions, confirmations, spending limits) come first. No promises of safety, security, compliance or specific behavior.

New to prompts? Start with the free ChatGPT Starter Prompt Pack: https://payhip.com/b/mxaFe

More packs in the store: https://payhip.com/PromptPackShop

Not legal, financial, HR or security advice.

Questions or a problem with your download? Email digitalpromohub.support@gmail.com.