Every guide to personalizing an AI assistant assigns the same homework: a twenty-question interview prompt, an identity profile, an about-me file to maintain. The most practical way to let your assistant actually know you is to skip that homework and connect it to the records your life already generates (calendar, sleep, workouts, location, messages) through a live personal context layer like Fulcra (which I co-founded). A questionnaire is only accurate the day you fill it out and decays from there, while a data stream is least useful on day one and compounds from there.
I spent twenty years in security, and the foundational lesson of that field applies directly here: identity claims are cheap, while behavior is expensive to fake. Bot detection does not work by asking the visitor whether it is a bot; it works by watching what the visitor does. When you personalize an assistant with an interview prompt, you are building the whole system on self-attestation, which is precisely the class of signal my old industry learned to discount first.
The research on human self-report justifies the discount. A meta-analysis of 106 effect sizes found that self-reported digital media use correlates only moderately with device logs and is rarely an accurate reflection of them (Parry et al., 2021, Nature Human Behaviour), and a systematic review of 187 studies found the same pattern for physical activity, with self-report running both higher and lower than what accelerometers measured (Prince et al., 2008, IJBNPA). You are not lying on the questionnaire; you are estimating, and humans estimate their own behavior badly.
Even the parts you get right start rotting immediately. The profile that says you are training for a marathon keeps shaping suggestions months after the race, because stale context does not announce itself. It just sits there being confidently wrong.
My working frame for personal computing is that a part of your life becomes useful to a computer once it is Observable, and everything downstream (computation, control, automation) inherits its quality from that first step. The observables that carry identity-grade signal are the boring ones you already produce: your calendar shows how you actually allocate time rather than how you would describe allocating it, your wearable knows what your sleep and recovery are doing this week, your location history knows your real routines, and your messages know what you are working on and with whom. Every one of these answers a question you would answer wrong on a form, and answers it fresh daily without you maintaining anything. And for the parts of your life no device observes yet (mood, habits, anything you would rather log than estimate), your agent can create the stream itself with Fulcra's custom tracking skill, schema, first entry, and dashboard included.
Some self-report genuinely belongs in writing, so here is the sorting rule I would use, and it requires nothing but a pen. I call it the Half-Life Sort. List everything your assistant ought to know about you, and next to each item, estimate its half-life: how long until that fact is more likely wrong than right. Your values, your food allergies, your kids' names, your hard constraints have half-lives measured in years, so write them down once in a context file and self-report will serve you fine. Your schedule, your recovery state, your current projects, your whereabouts have half-lives measured in days or hours, so writing them down is malpractice; those get a stream or they get nothing. The questionnaire approach fails because it stuffs short-half-life facts into a long-half-life container.
1. Write down the long-half-life facts. List what stays true for years (values, allergies, constraints, the people in your life) and paste it into your assistant's custom instructions. This is the only questionnaire you ever fill out.
2. Connect your streams once. Create an account at fulcradynamics.com and link your calendar, wearable, and phone, or install the getting-started skill and let your agent handle the setup itself.
3. Point your assistant at mcp.fulcradynamics.com. Any MCP-capable assistant can use it; test it with a question only real data can answer, like "when was my last full rest day?"
Everything after step three is compounding.
Is it safe to give an AI assistant access to my personal data? It is as safe as your scoping discipline. Grant each agent the minimum streams its job requires, prefer systems with per-agent permissions, and insist on being able to audit what any agent can currently see.
Do I still need an about-me file or custom instructions? Yes, for the long-half-life facts: values, constraints, allergies, relationships. The Half-Life Sort tells you which facts those are; everything else belongs on a stream.
What is the fastest way to connect my assistant to my real data? A hosted context layer like Fulcra is the fastest: connect your sources once and point any MCP-capable agent at the endpoint. The self-hosted route works too, if you are willing to be your own data engineer.
Fulcra was designed by people who get privacy and know the importance of an infrastructure solution that can be the secure private datastore for the rest of your life. Here data is yours, under your control, and only shared with the people and tools you choose to share it with.
.png){kind=logo}
