Questions to Ask Your AI
Twenty questions for the assistant you already use every day, rather than one you are trying out for the first time. They cover what it has stored about you, what your saved instructions are actually doing to its answers, when a thread has gone stale enough to abandon, and where you have quietly stopped checking its work. For getting a single piece of work right, Questions to Ask AI covers framing the request and pressure testing what comes back.
The questions
Open any question for the note
What do you already know about me before I say anything in this conversation?
Why ask it
Start here, because the answer varies by product, by account setting and by whether you are signed in. What comes back is often a mix of stored facts and inference, and seeing the inference is the useful half: it is what has been shaping answers you thought were neutral.
Read my custom instructions back to me exactly as you have them.
Why ask it
People write instructions once and forget them for a year. A playback catches the line you added during a single odd week that has been quietly bending every answer since, and it takes ten seconds to check.
Which of my saved preferences are you applying to this task, and which are you setting aside?
Why ask it
Instructions written for one kind of work often make another kind worse: a rule about brevity ruining an explanation, a tone rule ruining a formal letter. Asking which are live lets you suspend one for a single task rather than deleting it.
Is anything you have stored about me out of date now?
Why ask it
Jobs, cities, projects and relationships change faster than saved memory does. An assistant still writing for the role you left in the spring will produce answers that are subtly aimed at the wrong life.
How would you describe me to another assistant that has never met me?
Why ask it
A summary is more revealing than a list, because it shows what has been weighted. If the description is flattering to the point of uselessness, that is worth knowing about everything else it tells you about your work.
What have you retained that I probably did not mean to save?
Why ask it
A one-off mention of a health matter, a colleague by name, a client detail. Ask periodically and clear what you find, and check the product's own memory settings as well, since the assistant may not describe them accurately.
Where do I change what you store, and what does deleting it actually do?
Why ask it
Ask, then verify in the settings rather than taking the answer as authoritative. A model describing its own data handling is generating a plausible account, and the binding facts live in the provider's terms and your account controls.
Which of my standing instructions contradict each other?
Why ask it
Rules accumulate and start fighting: be concise against always show your working, be direct against soften bad news. When two conflict, one is being silently dropped, and it will not be the one you would have chosen.
Which of my instructions is making your answers worse?
Why ask it
You have to ask directly, because nothing volunteers this. Common offenders are rules about length, bans on caveats that suppress genuine uncertainty, and personas that make it agreeable when you needed it accurate.
What in this thread are you no longer applying from the beginning of it?
Why ask it
Constraints set early fade as a conversation grows, and the drift is gradual enough to miss. If the answer names anything at all, restate the requirements or start again with a clean summary.
Has this conversation got long enough that we should start again?
Why ask it
Long threads accumulate corrections, abandoned directions and contradictory instructions, and output degrades without any single moment where it went wrong. Asking prompts a useful assessment even though it cannot see its own context perfectly.
Write the summary I would paste into a fresh thread to carry on from here.
Why ask it
The move that makes starting over cheap. A good summary lists the task, the constraints, what has been ruled out and where you got to, which is also a fair test of whether it understood any of it.
What can you not see today that you could usually see?
Why ask it
Attachments, browsing, a connected account, files, the current date. Capability varies by plan, by setting and by the day, and getting the gap stated stops you asking for work that will be filled with invention instead of refused.
What do I ask you most often?
Why ask it
The answer is a fair description of what you have outsourced. It is worth looking at once a quarter, because the pattern is usually not the one you would have guessed and often points at something worth automating or learning properly.
What am I regularly using you for that you are not good at?
Why ask it
Names, numbers, current facts, anything with legal or medical consequence and anything requiring accountability sit badly with a tool that cannot be held to anything. A reply that claims to be fine at all of it is itself the answer.
What do I keep asking you that I would be better off learning to do myself?
Why ask it
Some delegation is sensible and some is quietly deskilling. The distinction usually turns on whether you would notice if the output were wrong, which is worth checking against your own last few requests.
Where have I been accepting your first draft when I should have pushed back?
Why ask it
Expect it to be agreeable here too, so treat the answer as a prompt rather than a verdict. Then go back through the last week and find the piece you sent on without reading properly, because there usually is one.
What have I stopped checking that I used to check?
Why ask it
Verification habits decay quietly as trust builds, and trust in a tool like this should not compound. If you cannot remember the last time you looked up a figure it gave you, that is the answer.
If I stopped using you tomorrow, what would get harder and what would get better?
Why ask it
The second half is the one nobody asks. A genuine answer might mention slower but sharper first drafts, or fewer confident wrong turns, and it is a fair way to size what the tool is actually worth to you.
What could I be asking you to do that I have never thought to ask?
Why ask it
Ask it having already told it what you do all week, or the suggestions will be generic. Ignore anything that reads like a feature list and keep the one or two that name a task you actually recognise.
Keeping a daily assistant in good order
Practical guidance for the conversation itself
Memory and settings
Read your custom instructions once a quarter
They are written in a hurry and then applied to everything for months. Reading them cold, with fresh eyes, is usually enough to find one line that made sense for a single task and has been degrading answers ever since.
Check settings rather than asking
What an assistant says about its own memory, retention and training use is generated text, not a readout. Ask if you like, then confirm in the account controls and the provider's terms, which are the only binding version.
Keep the personal and the professional apart
Separate accounts, projects or workspaces stop a client detail from turning up in a birthday message and a medical query from shaping a work summary. It also makes clearing memory a much less nervous operation.
Thread hygiene
Start fresh more often than feels necessary
Corrections pile up, early constraints stop being applied and the output drifts without any obvious failure. Asking for a carry-over summary and opening a clean thread costs a minute and resets all of it.
Restate the constraint that matters, out loud, late in the thread
If one requirement is doing the real work, say it again before the final draft. Repetition is cheap and the alternative is a polished answer that quietly stopped obeying it four messages ago.
Keep the prompts that worked
A note of the phrasings that produced good results is more useful than any general technique, because it is tuned to your work. Retry them after a product update, since behaviour changes between versions.
Where daily use goes wrong
Trust that compounds
A hundred good answers create a habit of not checking, and the hundred and first is where it costs you. Verification should stay tied to the stakes of the task rather than to your accumulated impression of the tool.
Instructions that make it agree with you
Personas and tone rules written for pleasantness suppress the pushback that made the assistant useful. If it has not disagreed with you in a month, look at your instructions before you congratulate yourself.
Treating a familiar tool as a private one
Daily use makes a text box feel like a notebook. Client data, health records, credentials and anything covered by an agreement deserve the same check on the hundredth day as on the first.
Asking it to grade its own work
Its assessment of whether it helped you is generated the same way as everything else. Judge it on your own outputs over a few weeks rather than on its account of them.