making space  /  ai playground  /  keep ai honest

How to keep AI honest.

AI tools sound sure whether or not they are right. They will say a job is done when it is half done, quote a number nobody measured, and forget the instruction you gave them an hour ago. These are eight habits I use to catch that. No code, no setup. Start with whatever tools you already use.

Which of these do you already do?

Tick the ones that are true today. You will get one next step. Nothing is saved or sent: close the page and it is gone.

The eight habits

Small, and doable this week.

Each one has the idea, what it looks like in ChatGPT, Claude or Notion, and one first move that takes ten minutes or less.

Habit 1

Write down what good looks like.

The AI can’t hit a target you have not described, and you can’t judge the result without one.

What it looks like

A short note in your custom instructions, a Claude Project or a Notion page: who the reader is, how long it should be, the tone, what a finished piece includes, and one example you love.

Your first move

Take the last thing you were happy with. Write five lines on why it worked, and paste them at the top of your next request.

Habit 2

Count what can be counted.

“Looks about right” is a feeling. A count is a fact.

What it looks like

Instead of asking whether a draft is too long, ask for the word count. Instead of asking whether it covered everything, ask for your checklist with a tick or a cross beside each item.

Your first move

Pick one quality you currently judge by feel. Ask for its count at the end of every request this week.

Habit 3

Enforce a rule, or stop claiming it.

A rule that nothing checks is a wish. Put something behind it, or stop saying it always holds.

What it looks like

You told the AI never to name a client in public posts. Nothing stops it. So you add a search for client names before anything goes out, or a “names removed” box in Notion you have to tick. Or you say honestly: “I usually remember to check.”

Your first move

List your three most important rules. Beside each, write what would notice a break. Where the answer is nothing, add a check or soften the rule.

Habit 4

Make the refusal say how to fix it.

A “no” that doesn’t say what to do next just gets worked around.

What it looks like

Tell the AI that when something fails, it names what failed and the exact next step. In a Notion form or template, the error reads “Add the client’s name in the Client field,” not “Invalid entry.”

Your first move

Add one line to your instructions: “If you can’t complete this, tell me what stopped you and what I should do next. Don’t guess.”

Habit 5

The one who does the work does not approve it.

Anyone reviewing their own work tends to agree with it. That is true of people and of AI.

What it looks like

Draft in one conversation. Review in a fresh one, or in a different tool, that sees only the finished piece and your note from habit 1. Or ask a colleague for two minutes.

Your first move

Open a new conversation. Paste only the result and your “what good looks like” note, and ask: “What is wrong with this? Be specific.”

Habit 6

Numbers come from something that measures them.

If nothing measured a number, it is a guess in a confident voice.

What it looks like

Figures taken from your spreadsheet, your accounting tool or a page you can open, with the source named beside them. Not “about 40 percent” with no trail behind it.

Your first move

The next time the AI gives you a figure, ask: “Where does this come from? Show me.” If there is no source, replace the number or label it as an estimate.

Habit 7

Keep standing instructions short.

The longer the standing brief, the less of it gets followed.

What it looks like

Your custom instructions fit on one screen: what matters most, in a few lines. Anything longer lives in a separate document the AI reads only when it needs it.

Your first move

Read your current instructions out loud. Cut every line that has never changed an answer.

Habit 8

The step you can’t undo stays yours.

Let the AI draft, sort and suggest. You send, pay, publish and delete.

What it looks like

Email set to save drafts, not send them. Automations that stop at “ready for review.” A publish or payment button that only you press.

Your first move

Write down the steps in your work that can’t be undone. For each one, check that you are the one who presses the last button.

Where this comes from

I run these on my own work.

My website is built with AI help every week, so it is where I test all of this. Right now 20 automatic checks run each time I try to send a change out. Each one is built to block a change that breaks a rule. That number is read from the site’s own files when the site is built. I don’t type it.

I have not got all of it right, and I found that out only by going back and counting. The details page says which parts work, which don’t, and what the mistakes cost.

The details are for peers I have invited. If you are on the list, you will be asked to sign in with your email.

Open the details Not on the list? Ask me for an invite.

Reviewed . I check this page every month.

Pick one habit and try it this week.

You don’t need all eight. You need the one you would not have ticked.

Back to the AI Playground