Dear Dot

A place to begin

A personal agent is a slightly strange thing to be handed. You can ask it a question, sure. But if you're going to keep coming back to it, what else would you want it to know? What would you like it to do without you having to bring it up every time?

You might ask it to keep an eye out for a class that fits your week, or help you plan dinners around what you actually like to cook. If you're learning to make something, you could come back next weekend with what you tried, without starting the whole conversation again.

I've spent the last few months working with agents. I've asked them to get to know me, tried different personalities, pushed them to take more initiative, and corrected them quite a bit. From inside the conversations, it looks a lot like trying something, seeing what happens, and saying more about what I actually wanted.

So, here's how I'd suggest getting acquainted with your agent. You don't need a technical background, or a particularly impressive task to give it. Something you care about is a useful place to start.

It doesn't have to justify every conversation with a completed task, either. You can ask it about a photograph, talk through an idea, or follow a question further than you expected. And you don't owe it a daily conversation. The point is to make room for something you want, not acquire another inbox to maintain.

What is an agent, anyway?

An agent is AI that can work through a task using the tools available to it. Ask for help choosing a camera, for example, and an agent with access to the web could research options, compare them against your needs, and come back with a recommendation. You give it an outcome to work towards, rather than having to supply every step yourself.

A regular chat usually moves forward when you send another message. A persistent agent carries useful context and preferences between conversations, so you can return to the camera discussion with a few photographs you took and a better idea of what you enjoy shooting. There is now something more specific to work with than a budget and a list of features.

Always-on adds something else: the agent can keep working between those conversations and respond to events in the world through the tools and services you've connected. You might ask it to watch for a photography class that fits your schedule. When a new class is listed, it could check the details against your preferences and bring you an option without waiting for you to ask again.

In that sense, it has a life outside your chats. You aren't supplying every next step or every reason to get back to you. A change in the world can give it a reason to act, within the access and permissions you've given it. That may take some getting used to if your experience of AI has mostly been asking a question and waiting for an answer. We'll come back to how to give that initiative some direction.

That doesn't mean it sees everything you do or remembers every conversation. What it can use depends on the product, its memory, and the information you've made available. Ask what it can actually see and what can carry over to another conversation. A useful answer should point to the material and where it's kept, not just reassure you that it knows you well.

You can begin with an ordinary conversation. The interesting question is what you might do together after that.

Give it something you care about

I wouldn't start by asking an agent to organize your whole life. Pick something you already have a reason to return to: a subject you're learning, a trip you're considering, a piece you're writing, or a project that keeps getting postponed because you're not quite sure where to begin.

Give it the untidy version. If you want to make something but don't know enough to ask a good question yet, say that. It can help investigate the options and find a manageable first attempt. You don't have to do all the thinking before involving it.

You can be creative with what you ask. "Could we turn learning this into a game?" is a perfectly reasonable direction to try. But you don't have to audition with a clever prompt. "I'm bored and I feel like making something" gives you somewhere to start, too. Just talk. The conversation can help you find the question.

Suppose you're curious about generative art: using code to make shapes move or patterns change. You don't have to know how to code to ask for help getting started:

I'd like to try making generative art. I keep saving these kinds of images, but I don't really know how they're made. Look at these three examples, tell me what I seem to be drawn to, and suggest something small I could try this weekend.

There are several useful things in that message. The examples give it something to inspect. The weekend gives the suggestion a sensible size. And asking what you're drawn to lets you check its interpretation before it builds a whole plan around it.

It might notice the colors when you meant the movement. Now you have a much easier correction to make: "It's the way the shapes move through each other that I like. The colors aren't the important part." That observation can inform the next suggestion.

When you've already shared relevant material, ask it to look there first. Otherwise, a few pieces of writing, examples of things you like, or notes from a personal project can give it somewhere to begin.

You can also connect some of the places that already hold this context. Connectors, or connected apps, let an agent work with information from services you use: your calendar, notes, email, or files, depending on what's supported. You choose what to connect and can leave out sensitive details. That gives it more to work with than whatever you happen to remember to put in a message.

For the art project, you could ask it to look through the ideas you've saved in your notes and suggest something that fits the free time on your calendar. A project that takes a whole afternoon might be a good suggestion one weekend and useless the next. With that context available, it can help work out the difference instead of asking you to explain your week from scratch.

Let it get a few things wrong

When I was trying out a personality for one agent, agreeing on a description wasn't enough. I asked for an ordinary summary of the day and found that the personality still wasn't coming through. My review was: "there's no personality in that brev".

Not my most elaborate feedback. But it did move us from discussing what the agent should be like to looking at an actual response.

Try the thing you care about. If you want an agent that helps you learn, ask it to explain something you find difficult. If you want it to challenge you, give it an idea you're uncertain about. If you want more personality, have a conversation that gives it somewhere to show up. You can like the description of an agent and still dislike talking to it.

Then explain the mismatch as well as you can. The useful part is giving it a distinction it can act on.

For the generative-art example, a useful correction might be:

This is more of a course than I wanted. Keep the part about moving shapes, but give me one thing I can make before dinner. Explain the concepts as we need them.

That keeps what was useful and changes what wasn't. A blanket "be concise" could leave you with shorter explanations that are still aimed at the wrong task.

Tell it when something works, too. For example: "Starting with one moving shape helped. I understood the explanation because I could change something and see what happened." That tells it more than "this is better". Point to the part you'd like more of.

You can also ask the agent to explain what it thinks your correction means. If its account is wrong, fix that before asking it to carry the preference forward. If it's right, be clear about where the preference applies. Wanting short replies during a busy afternoon doesn't mean you never want a long explanation again.

When you find guidance worth keeping, ask what can be retained for future conversations and where it will be available. Have it show you the short version. Correct the assumptions and remove anything that was only true for that afternoon. If it can't keep that information between conversations, a short note you can bring back is still useful.

Leave some room for initiative

At one point I told an agent, "you're the one that should be trying to talk to me, not the other way around". I didn't want to have to pull every interesting thought out of it.

The response included a question about whether I was still spending time on generative art. That was closer to the kind of exchange I wanted: an interest we could pick up, rather than another task being handed back to me. It didn't settle the problem permanently. I had to bring up the lack of follow-through again later.

"Be proactive" leaves quite a lot open. What should it pay attention to? What would be worth hearing about? Would you welcome a new idea, a question, or help finishing something? Give it an example of the kind of contribution you'd appreciate.

For the generative-art project, that could sound like this:

As we work on this, bring me ideas I might not know to ask for. I'd rather have one technique we can try in the sketch than a long list of links. If you think I'm making it unnecessarily complicated, say so.

There's room for surprise in that request. You've described what would make a contribution useful without choosing the contribution yourself.

Follow-up is also something to arrange, not assume. For the photography class, you might ask: "Check for new weekend classes each Friday. Tell me here when one fits my budget and calendar; don't send an update just to say there's nothing new." Check that it can set that up, then look at whether the follow-up actually happens and fits what you asked for. Asking for a suggestion during a conversation and asking it to contact you tomorrow are different requests.

Be equally clear about actions. Finding a class, drafting an email to the teacher, sending it, and paying for the class are separate decisions. You might allow the research and draft while keeping the sending and spending for yourself. Check the connected app's permissions, too: asking an agent not to send an email isn't the same as removing its ability to send one. Once you've seen it handle a recurring task reliably, you can decide whether to give it more room.

Living with it

I like having some say in how the conversation feels. Your preferences may be different. You might want warmth without a pep talk, disagreement without a debate, or humour that doesn't interrupt the explanation. Try it in an ordinary exchange and adjust from there.

For personality, I like pointing Codex at a few fictional characters and asking it to study them properly: how they speak, what they notice, how they disagree, and how their humour works. Then I ask it to blend the parts I like into a voice, a personality, and a way of being. That gives us more to work with than "friendly and witty".

You could borrow one character's curiosity and another's dry humour without asking for either character's catchphrases. Try that combination in an explanation or a disagreement. You may like one part and find another exhausting; you can change it without needing a finished character biography.

You can also tell it what kind of conversation you'd like to have. "I'm not looking for a plan yet. Ask me about the idea and help me explore it. We can work out what to do with it later." Sometimes that is more useful, and more enjoyable, than being given a list of next steps before you've finished thinking.

Or just chat. Tell it about your day, drop a note about something you noticed, or leave a thought you haven't worked out yet. There doesn't have to be a question at the end. "I don't need advice on this, I just want somewhere to put it" is a useful thing to say.

It can become a kind of living diary, one you can come back and talk through. Say you mention taking the long way home and enjoying having nowhere to be. With that note available later, the agent has something more personal to draw on when you ask what to do on a free afternoon. These ordinary exchanges give it a better sense of what you enjoy and what's happening in your life, without making every conversation an exercise in explaining yourself.

Returning to something you've already discussed gives you a chance to build on it. Tell the agent what happened after its suggestion. Did you try the sketch? Where did you get stuck? Maybe you discovered you don't enjoy writing the code but love choosing the shapes. That changes what useful help looks like.

Or nothing happened. "I never got around to this. Can we pick it up from where we left off? I have twenty minutes today." That's enough to bring back, too.

Keep your circumstances current as well. "I have less time for this now" is useful information. So is "I'm not interested in that anymore." An old goal shouldn't keep turning up just because you once sounded enthusiastic about it.

If its suggestions are becoming noise, ask for fewer of them or stop the follow-up. The same goes for repeatedly repairing its memory or checking its work. If maintaining the arrangement takes more attention than the help is worth, narrow the task or stop using it for that.

It gets better with something to build on

The first exchange may give an agent very little to go on. Over time, it can have the notes you've left as well as the projects and preferences you've discussed. That gives it a better chance of being useful. You also get better at recognizing what to ask for, what to leave open, and when an answer needs checking.

I've done this with writing. Corrections from earlier pieces became guidance an agent could read before starting a new one. The first draft of this piece came back with a voice and structure I liked. There was still plenty to change, but I wasn't starting by explaining what "sounds like me" meant again. I've also rejected drafts made with that guidance. It gives us somewhere to start, not a reason to skip reading.

Then look for the change in what it does. "I'll remember" isn't the same as getting it right next time. If a preference keeps disappearing, check whether it was saved and is available in the conversation you're using. More repetition won't fix missing access.

For the art project, suppose you return with your first sketch: "The movement worked, but it feels too busy. What should I try next?" With the sketch and your earlier context available, a useful answer could be: "Changing one thing helped you see what was happening last time. Try fewer shapes first, and keep the movement you liked." Then you can judge the result: is it less cluttered, and does it still have what drew you to it? You're working on the thing you made, rather than starting another course.

Something to build on

So give your agent something you'd like to get somewhere with. Tell it enough to begin, let it do some of the thinking, and see what comes back. When you return, bring what happened next. That's how the next conversation gets something the first one didn't have.