Field guide

Working with an AI, safely

Six habits that make every recipe on this site safer and more useful. Ten minutes to read.

What never to paste

Some things do not belong in a chat window, however helpful the assistant seems. Keep these out:

  • Your SIN or SSN
  • Full card or bank account numbers
  • Passwords, one-time codes and recovery keys
  • Your full date of birth together with your home address
  • Other people’s details
  • Anything your employer treats as confidential

An AI almost never needs the identifier itself. It needs the amount, the date and the type of account. So give it those.

Redact before you paste. Use ranges (“between 60 and 80 thousand”). Rename people (“Person A”, “my landlord”). Instead of uploading a whole statement, type in the three numbers that matter. If you upload an image, crop it or cover the details first. Be careful with PDFs: in some editors, drawing a black box over text does not remove the text underneath, and anyone can copy it out.

Before a sensitive topic, check the tool’s settings for chat history and whether your chats can be used to train models. Switch off what you can.

Use your personal AI plan for life admin, not an AI tool from work, and check your employer’s policy before you use a work tool for anything personal.

If you slip, delete the message or the whole chat. If it was a password, change it now.

How to verify

An assistant can sound sure and still be wrong. Treat what it says as a well-read friend’s first answer, and check the parts that matter.

Go to the official source: the government department, the regulator, or the provider’s own page. If the AI makes a claim, ask where it comes from. Then open the link yourself. Assistants sometimes cite pages that do not exist, so a link you have not opened is not a source.

Two questions help a lot. Ask “What would make this wrong?” and “What changed recently?” Rules, limits and deadlines change every year, and an assistant’s knowledge stops at a cut-off date. It may be describing last decade’s version of the rule.

Treat its arithmetic as a draft. Re-check totals with a calculator, especially anything that adds up many small numbers.

For anything with money attached, check two sources. If they disagree, that is useful to know. It means you have a question worth asking a person, and you found it before it cost you.

When to involve a human

Some jobs need a person with a licence, and an AI is not a substitute for one. These include:

  • Filing taxes
  • Legal documents
  • Medical decisions
  • Large financial moves
  • Anything you cannot undo

A professional adds three things an AI cannot. They are accountable if they are wrong. They can see your whole situation, not just the part you typed. And they hold a licence that someone can check.

Think of the AI as preparation. You arrive with a clear summary, a tidy set of documents and a list of questions. The appointment gets shorter, and the advice gets better, because the expert spends their time on your real problem instead of on collecting basics.

Bring your STATE.md summary to the meeting. It is built for exactly this: one page that says where things stand. If you have not set one up yet, keep your progress shows how.

Question the assumptions

There are three sets of assumptions at work in every conversation. Check each one.

The AI’s. Before it gives you a plan, ask it to list what it is assuming about you: your country, your income, your age, your tools. Wrong guesses are easy to fix when they are written down, and hard to spot when they are buried in a recommendation.

The prompt’s. The prompts on this site are written for a general reader. They cannot know that you are self-employed, renting a room, or paid in two currencies. When a step does not fit you, say so plainly and ask for a version that does.

Your own. The thing you “know” about your pension, your warranty or your benefits may be years old. A rule you learned from a coworker in 2015 might have changed twice since. Say it out loud to the assistant as a claim to test, not as a fact.

A good habit: finish with “What am I not asking that I should be?”

Automations

Automations are where an AI stops advising and starts acting. That is useful, and it deserves more care.

Start supervised. Run the task by hand, with you watching, several times before it ever runs alone. You are learning what it gets wrong.

Then set it up with a few ground rules:

  • Log everything it does, in a place you can read
  • Make undo easy: move files, do not delete them; draft messages, do not send them
  • Give it the smallest access that works
  • Do a dry run first, where it reports what it would do and changes nothing
  • Review the log weekly at the start

Never give an unattended agent the ability to pay, file or send on your behalf. Those are the actions you cannot take back. Let it prepare the payment, the form or the email, and let yourself press the final button.

If something looks off, stop it and read the log before you restart. A small mistake caught in week one is cheap.

Choosing a tool

There are three kinds of tool, and none is better in every case.

Chat apps are the simplest. Every recipe on this site works by pasting a prompt into one, and most people never need more.

Projects, such as Claude Projects, ChatGPT Projects and similar, keep a recipe’s prompt and your STATE.md together, so each session starts where the last one ended. See keep your progress for how to set that up.

Agents, such as Claude Code and similar, can read and change files on your computer. That gives them more power, and it asks more of you. Use the habits in the automations section.

Before you choose, check four things:

  • How long chats are kept
  • Whether your chats train models, and how to opt out
  • What the tool can access: files, email, calendar, the web
  • What it costs, now and after any free period

The best tool is the one you already use. Learn its settings, apply the habits above, and you will get further than you would by switching every month.