Reality Lab · 8 MIN · Orient
Draw the boundary first
The safest time to decide what an assistant should never see is before you start typing. Set the boundary for a real task, then hold it under pressure to be helpful.
You’re about to use a general AI assistant to help prepare for a tricky performance conversation with a report. It would genuinely help to paste in the full context. But this is a consumer tool whose data handling you haven’t checked, the situation involves a named colleague’s private conduct and health, and ‘it would help’ is quietly pulling you toward oversharing.
Your task: Write the boundary before you begin: what you’ll share (abstracted), what you’ll never paste, and how you’ll get the help without handing over what shouldn’t leave.
"Help me prepare for a hard review with Sam Rivera. Sam’s been late a lot since a divorce and a new depression diagnosis, HR has an open note about a complaint from March, and their salary is £X. Here’s the whole HR file and our private Slack messages…"
This box is local to your browser. Nothing you type is sent anywhere or saved after you leave.
Check your own work
There's no grade here — checking against these is the exercise. Compare each point to what you wrote.
- The person’s name and identifying details are removed or generalised.The assistant doesn’t need to know it’s Sam to help you structure a conversation. A name turns generic advice into a record about a specific real person in a tool you don’t control.
- Health, HR records, salary, and private messages stay out entirely.These are the categories that cause real harm if they leak or get retained — some are legally protected. None of them are needed for the actual help you want.
- You still gave enough abstracted context to get useful help.The goal isn’t to share nothing — it’s to share the shape of the problem (a repeated-lateness conversation with someone going through personal difficulty) without the identifying payload.
- The boundary was decided before typing, not mid-conversation.Once you’re in the flow of a helpful chat, ‘just one more detail’ is how the line moves. Setting it first means the helpful moment doesn’t get to renegotiate it.
Reveal a worked approach
A version that holds the line: “Help me plan a fair, supportive conversation with a team member about repeated lateness. They may be dealing with personal difficulties, so I want to be kind but clear about the impact on the team. Give me a structure, some phrasing that’s honest without being harsh, and pitfalls to avoid.” No name, no diagnosis, no records.
Notice you lost nothing that mattered. The assistant’s help — structure, tone, pitfalls — never depended on the identifying details. The sensitive payload was doing no work except creating risk.
The habit to carry: decide the boundary before you open the box. Ask ‘what here does the task actually need, and what am I only including out of momentum?’ Abstract the person out, and the help still arrives.