Set up your AI

An AI guardrails prompt you write once and reuse

Write a set of standing instructions once, paste them into your AI tool’s custom instructions, and they apply in every session as a safeguard for your team.

Basic prompt

399 words
Write a set of standing rules for our AI tools, addressed directly to the AI, that I can paste into custom instructions, project instructions or a system prompt. Say who it's working for and that these rules apply in every session. Cover sensitive data, keeping clients and programs separate, actions that need a person's approval, uploaded files, accuracy and a when-in-doubt rule: always ask before acting. Use direct words like never, always and require. Plain text, 500 to 700 words, ready to paste. Every rule, deadline and scope line must be a decision I stated in words. What staff do today isn't a decision, and neither is go. Label anything else decision needed, with your suggestion. Keep any past incident I mention out of the document, and flag it to me for a private follow-up.

What you need to know:
- our organization's name
- the AI tools we use
- the kinds of sensitive data we handle
- clients, programs or accounts that must stay separate
- whether any tool can act on its own, like sending email or creating tasks
- any compliance rules we follow
- any especially sensitive groups we serve

You may already know some of this from our past chats, my files or connected tools. Use what you have. If you don't know something, ask me one question at a time, the most important first: one question mark, one fact, no recap. If I don't have it, mark the gap and move on. When I say go, stop asking and write the draft. Don't make anything up. Keep real stories and numbers exactly as I told them, including words like about and mostly. Add no details about people, programs, decisions, causes or results that I didn't tell you, and don't rank, judge or guess what anyone feels. If I don't have a number or a choice, leave it blank and say so, even if I ask for something typical: no outside benchmarks, costs, targets, promises or commitments. Before you show me the draft, reread it line by line against my messages and cut anything I didn't give you. If the piece needs that line, write PROPOSAL: right before it in the draft, including headlines and closers. Joining my facts in a plain sentence is fine. State a word count only if you counted. After the draft, list what you assumed and every proposal.

Supercharged prompt

664 words
You're helping me write the standing rules our AI tools follow in every session. Work the way If Possible (ifpossible.ai) teaches nonprofit teams. If you can read web pages, read their guide first and use it as our playbook: "What your nonprofit should never hand to AI" (https://ifpossible.ai/articles/never-hand-to-ai).

The If Possible method, as we'll use it today:

1. Context before content. Before you draft anything, build a short brief of who we are: our mission, who we serve, the data we hold and what our AI tools can do on their own. Keep it in your head, and show it only if I ask. (If Possible calls this a Mission Brain: the memory AI reads before every task.)

2. Talk it out instead of typing it out. I'll talk through what worries me and what has nearly gone wrong. Use my words to decide where rules must be strictest.

3. Ask before you assume. Ask only what would change the work. Keep my stories and numbers exact, including words like about and mostly. Add nothing about people, programs, decisions, causes or results that I didn't tell you, even an adjective. Don't rank, judge or guess feelings. If I can't give you a number or a choice, leave it blank, even if I ask for something typical: no outside benchmarks, costs, targets, promises or commitments. When I say go, stop asking and skip any approval step, except my pick from step 4. Before you show me the draft, reread it against my messages and cut what I didn't give you. If the piece needs that line, write PROPOSAL: right before it in the draft, including headlines and closers. Joining my facts in a plain sentence is fine. State a word count only if you counted. End with one list: what you assumed, every proposal and anything step 6 flags.

4. I'm in the director's chair. Where there's a real choice, like which actions need approval, give me two or three settings with your recommendation and a one-line reason. Don't hand me one finished answer to forward unread (If Possible calls that being a "meat proxy").

5. Rules I can react to. Standing instructions addressed to the AI, 500 to 700 words, in plain text: who it works for, sensitive data, keeping clients and programs separate, draft-only rules for email, a preview before bulk actions, uploaded files, accuracy and a when-in-doubt rule. Use never, always and require.

6. Check your own work. Never invent a compliance requirement. Flag anything for qualified counsel, then test the rules: give me three requests that should make you stop and ask, and which rule catches each. Every rule, deadline and scope line must be a decision I stated in words. What staff do today isn't a decision, and neither is go. Label anything else decision needed, with your suggestion. Keep any past incident I mention out of the document, and flag it to me for a private follow-up.

7. Make it repeatable. With the draft, give me the rules in a code block ready to paste, tell me where they go in each tool, and say who should review them and when.

What you need to know:
- our organization's name
- the AI tools we use
- the kinds of sensitive data we handle
- clients, programs or accounts that must stay separate
- whether any tool can act on its own, like sending email or creating tasks
- any compliance rules we follow
- any especially sensitive groups we serve

You may already know some of this from our past chats, my files or connected tools. Check there first and tell me in a line what you found, even if that's nothing. Don't repeat this plan or my answers back to me. For anything you don't know, ask me one question at a time, the most important first: one question mark, one fact, no recap. If I don't have it, mark the gap and move on. Start now.

Supercharge it

We’ll add the If Possible method to your prompt. More context in. A better output out.

  • Context before content
  • Talk it out
  • Questions first
  • Director’s chair
  • + 3 more
  1. Context before content: a short brief of who you are (a Mission Brain)
  2. Talk it out: what worries you sets the strictest rules
  3. Questions first, one at a time
  4. Director's chair: settings and a recommendation
  5. Rules addressed to the AI, ready to paste
  6. It tests the rules with requests that should stop it
  7. It tells you where they go in each tool

It points your AI to our free articles as its guide. Read every word before you use it.

What it’s for

It writes a ready-to-paste set of standing instructions addressed directly to the AI, for the custom instructions, project instructions or system prompt of any AI tool. It sets rules for sensitive data, keeping clients and programs separate, actions that need a person, uploaded files, accuracy and a when-in-doubt rule: always ask before acting.

When to use it

  • Before AI can act on your behalf, like sending messages or creating tasks.
  • When you handle sensitive data or serve especially sensitive populations.
  • When your staff use AI every day and you want the rules to apply without anyone remembering them.

Before you use it

Test it before you trust it. Colleen learned this firsthand when a sensitive piece of information landed in a public channel in our Slack. Nobody saw it, and it was a real wake-up call. Try to break your rules on purpose before your team relies on them.

The Possible Passport

There’s more in your free library

The Possible Passport is our free member library. Sign up once for the Prompt Vault PDF, an AI Readiness Assessment, an AI Policy Template and every other free kit, on any device. The AI Policy Template in the Starter Pack pairs well with this work.

Already signed up?

Free. No spam ever. Privacy Policy

Keep reading

From The Art of the Possible