The assistant I want isn't a chat box
I've had useful conversations with AI about finding my next job, renovating our house, buying a new car, and planning for retirement. In each of those conversations it feels like having someone thoughtful working alongside me. Then I close the window, and I'm still the person holding everything together.
I have to remember which projects need attention, notice when something changes, find the conversation where an important decision was made, and update the AI on what has happened since. The answers are great, but keeping my life moving forward is still my job.
That gap is what led me to build an assistant of my own: something that keeps the context, advances the work while I'm busy elsewhere, and brings me decisions it has already helped prepare, along with a place where I can see the work and steer it.
A chat box is a fine way to talk to my assistant, but I don't think it should be the place where the whole relationship lives.
What I want from an assistant
I use Claude and ChatGPT every day. Chat is an unusually flexible interface: I can arrive with a vague question, work through an idea, push back on an answer, and change direction halfway through. I want to keep all of that. But when I think about an assistant I'd trust with an ongoing role in my life, I want four more things.
- Know what's happening. What did we decide about the renovation? Did the recruiter reply? Which deadlines are real and which are made up? It should keep information current and maintain a record of our decisions that I can read and correct.
- Understand what matters to me. Knowing I'm looking for a job is a start. Knowing what I want from my next role, what I'll compromise on, and which attractive sounding roles might not actually be a great fit is what makes this uniquely helpful and useful.
- Recognize my blind spots. I like forward momentum. I can get very good at answering a question before considering whether it was the right question from the start. An assistant that knows me should catch that.
- Take initiative and follow through. Keep watching. Do the prep. Return only when there's a decision, a recommendation or a task that needs me, and then track that work to make sure it actually got done.
Taken together, that is a much bigger ambition than responding in a chat box. I want an assistant that uses what it knows about me to help me become the person I'm trying to be.
My life happens in parallel. A conversation doesn't.
That ambition changes what I need from the interface.
The job search doesn't pause while I'm dealing with housing permits. A financial decision can sit unresolved while a dozen newer conversations push it out of view. Family, health, money, work and friends stay important even when I haven't mentioned them lately. A conversation has an order. My responsibilities have relationships, deadlines and competing claims on my attention.
I need to open something and see the situation: what changed, what's progressing, what needs me, and what can wait. I want to click into my home renovation project, see that I need to book a site surveyor, peruse the 3-5 recommended firms to call along with key questions to ask, and once I've booked them, I want the assistant to continue moving the project forward.
That's why I'm building toward a dashboard rather than a thread. A dashboard doesn't create memory or initiative. Those come from connected information, a maintained record, and agents that run between conversations. What the dashboard gives me is a way to see and direct that work. Without it, I could have a capable system behind the scenes but still spend my mornings piecing together context.
Chat stays where I go to think out loud, explore something new, or explain a complicated change. It's the terminal now, not the home screen.
My first attempt
The Assistant Dashboard is a first attempt at this. It's one page that puts every project in a fixed place, prioritizes the items waiting on me, and gives me a way to read what the assistant has prepared and reply from the page. It's two months old and has been through about a hundred changes, and I don't yet love it.
I'm still learning what deserves space, what deserves urgency, and what should fade into the background until it matters. Early on, every overdue item turned red, including "deadlines" I'd casually set for myself, like "pick an estate attorney." Within a week, red meant nothing. Now only hard deadlines with real consequences, like a deposit due or a cancellation window closing, get to be red.
A few pieces already show me why this is worth pursuing.
Job Radar: learning what I value. Every morning it reads PM openings at a few dozen companies I'm interested in and scores each one against a written description of what I want to do next. The strongest fits reach me with a one-liner on why. I mark each one as Interested or Not for me, and sometimes tune the AI with a note.
The most useful part of building this was writing that description. "Find me an interesting product job" left too much unsaid. I had to write down what I was actually optimizing for: room to originate new ideas, how close the team is to AI, pay floors, hours, and the tradeoffs among them. That made my own thinking clearer, and my feedback now gives the system something better than a list of titles. It gives examples of how I decide.
Project Books: continuity beyond a conversation. Every project, from the home renovation to the job search to a What's for Dinner planner, has its own Notion page with the same four parts: a brief, a decision log, open questions, and a log of what's been done. An agent reads the book before it works and writes a line when it finishes. When I come back to a project, I open a fresh conversation, point it at the book, and the context has a durable home outside the chat.
This isn't perfect memory. A record can be incomplete or wrong. But I can read it, correct it, and see exactly what information the assistant is working from, and that matters as much as remembering.
Purchase Advisor: catching my blind spots. This is the example that feels most personal. I enjoy research. Comparing options feels productive, and once I start, the decision quietly becomes which one should I buy? I recently found myself deep in car research for a car that I realistically didn't need until next summer. Sometimes research gives me the illusion of progress.
So the advisor has a job before it compares anything: ask whether I need the thing, why now, and what happens if I wait. "Waiting" can be a useful recommendation, but when I'm deep in research and enjoying the process, I often don't even consider it. This is just one example of the assistance I want: something that knows a pattern in my behavior and helps me manage it.
Some projects are just for fun. The Concert Radar reads the venue newsletters I never open and flags artists I'd want to see before the presale starts so I buy at face value instead of being gouged at resale. Helping manage my life should also include the fun parts.
Initiative needs boundaries
Giving an assistant initiative means deciding where that initiative ends.
In an early version, recommendations sometimes showed up in a project's decision log as though I'd made them. Even when the recommendation was reasonable, the record was wrong, and it blurred the line between the assistant's view and my choice. That led to a clearer rule: it should prepare recommendations, but decisions or changes to preferences require my approval.
The same principle runs through the rest of the system. It drafts messages; I review and send them. It researches purchases; I decide whether to buy. It prepares recommendations; I make the calls. These are written instructions, not technical limitations, and trust grows as I watch how reliably they're followed. The full set is in my AI System.
A personal board of directors
During my MBA at Kellogg, I joined a program called Personal Board of Directors or "PBOD". I was matched with three other students, and we met biweekly for a year to be sounding boards for one another.
We could bring whatever was on our minds. For most of us, that meant careers: which internship to take, which job to pursue, what to do after business school. But as we got closer, the conversations widened. People shared personal struggles, family issues, financial concerns. We learned more about the individuals behind the career decisions.
Over time, you didn't have to explain all the context. The people around the table knew some of your history. They understood what mattered to you. They could bring a different perspective to a decision while keeping your values at the center of it.
It's been a decade since my MBA but that experience has stayed with me as I think about what I want from an AI assistant.
The relationships made it special, but the structure offers a useful model: continuity, different perspectives, a safe space to be gently challenged, and advice grounded in an understanding of your life.
I'd like to build toward a personal board of directors in that spirit. Each agent would pay attention to an area I care about, bring an informed perspective, and help me see how a decision fits into the larger picture. I've already had one-off conversations with AI about my health goals, our spending, retirement savings and the challenges of balancing parenthood while retaining my existing friendships. I want each of those to become a seat at the table:
- A health seat that tracks my bloodwork, DEXA scans, and wearable data over time, notices trends, and prepares questions before doctor's appointments.
- A finance seat that reviews spending and savings, updates retirement projections as circumstances change, and models how a decision like the renovation budget shifts the range of outcomes.
- A relationships seat that helps me follow through with the people I care about when work and family crowd everything else out.
Like a real board, they'd disagree. The finance seat wants me to save more while the relationships seat sees a chance for a trip with close friends. A career move improves our finances but leaves less time with my family. I'd want the board to make those tradeoffs visible, argue each side, but leave the final call to me.
Unlike the Kellogg version, this board never adjourns. It keeps paying attention to the things I've said that matter while I'm busy with everything else, and its ultimate goal is to help me build the life I want.
Something everyone could have
My system is a few early pieces of that puzzle, and making the whole thing reliable will take work. But using those pieces has changed what I expect an assistant to be and what I believe it can become.
Assembling one is also a different thing from using one. Most people shouldn't have to configure scheduled jobs, design a memory system, or build a dashboard to get this kind of help. Those are problems a product should solve.
What I want to open each morning is useful work already underway, a few decisions waiting for me, and a group of agents working on my behalf to advance the things in life I care about most.
If you're building something in this shape, or run a team that should be, I'd love to compare notes: tylersanchez@gmail.com.