AI That Helps You BE, Not Just DO

Table of content

by Ray Svitla

thursday, 14:32. I open Claude after two hours of coding.

it says nothing about the two hours. doesn’t notice I’m about to crash. doesn’t know that — based on 16 days of my own data — I lose focus like clockwork at exactly this point.

it just sits there, cheerful and blank: “how can I help you today?”

this is the most advanced AI on the planet. it can write poetry, debug code, explain quantum mechanics. but it has no idea I’m about to spiral into scattered app-switching because my brain just hit a wall.

I know this because I tracked it. manually. like some kind of digital anthropologist studying myself.

the do economy

here’s what every AI tool does right now: completes tasks.

┌─────────────────────────────────────┐
│           current AI                │
│                                     │
│    user asks  →  AI answers         │
│    task in    →  result out         │
│                                     │
│         (blank between uses)        │
└─────────────────────────────────────┘

ChatGPT writes your emails. Claude reviews your code. Cursor autocompletes your functions. give task, get result. they’re genuinely impressive.

and now they have memory. ChatGPT remembers you prefer dark mode. Claude remembers you’re building a startup.

but here’s what none of them do: notice.

they remember WHAT you told them. they store facts: your name, your project, your preferences. like a notebook that writes itself.

they don’t observe HOW you work. when you lose focus. what triggers your best thinking. why you’re more productive on thursdays than mondays.

a notebook is passive. it records what you put in.

a mirror shows you what you can’t see yourself.

every AI you use right now is a very smart notebook.

the missing layer

I’m not talking about productivity tracking. screen time reports exist. they’re useless — just guilt without insight.

and I’m not talking about Facebook-style prediction. they know your patterns better than you do, but they use that knowledge to keep you scrolling. the model wins when you lose.

I’m talking about something different:

        DO                          BE
        ──                          ──
  "write this email"         "you crash at 14:00"
  "fix this code"            "protect 11-12:30"
  "summarize this doc"       "you've drifted 3 days"
        ↓                           ↓
   task completion           self-knowledge

the BE layer.

not “what do you want me to do?” but “here’s what I noticed about how you work.”

it’s the difference between a tool and a coach. between efficiency and clarity.

what 16 days revealed

I tracked my activity for two weeks. here’s what I learned that no AI ever told me:

I warm up with YouTube every morning. 85% of days. 15-30 minutes. phone reviews, programming history, random business interviews. eight days in a row without breaking the pattern. I thought I “sometimes” watched stuff. turns out it’s a ritual.

I lose focus after exactly two hours. like clockwork. every deep work session ends the same way: 30 minutes of drift. Telegram, random tabs, scattered browsing. not a decision. a crash.

my best work window is 11:00-12:30. this one hurt. I often schedule calls there. ninety minutes of prime cognitive real estate, and I’ve been giving it away to “quick syncs.”

Telegram appears in 80% of all 30-minute blocks. it’s not a communication tool. it’s background static.

when I work past 22:00, my mornings drift. midnight monday → 10am start tuesday. three nights in a row and suddenly it’s noon before I hit deep work. obvious in data. invisible in the moment.

here’s one day, raw:

feb 6 — my day in patterns

09:16 ░░░░░░░░░░░░░░ youtube warm-up (14 min)
09:30 ████████████████████ research + planning (46 min)
10:16 ░░░░░░░░ telegram/slack cluster (29 min)
10:46 ████████████ deep work — rebranding (31 min)
11:17 ░░░░░░░░░░░░░░░░ scattered browsing (30 min) ← crash
11:47 ████████████████████████ hubspot editing (60 min)
12:47 ░░░░ gemini exploration (15 min)
13:02 ████████ google meet (45 min)
13:47 ░░░░░░░░ lunch break
14:17 ████████████████ client work (48 min)
15:05 ░░░░░░ slack/telegram (22 min)
15:27 ████████████████████ deep work — self.md (55 min)

░ = distraction/break     █ = focused work

you can see the 2-hour crash at 11:17. you can see the morning warm-up ritual. you can see focus fragmenting after lunch.

I have 3 years of ChatGPT history. hundreds of Claude conversations. not one ever said: “hey, you crash around 11:17” or “your best hours are 11:47-12:47.”

they remember what I told them. they noticed nothing about how I work.

why nobody builds this

the obvious question: why doesn’t it exist?

Meta and Google have your data but use it to sell ads. to keep you scrolling. their prediction models are designed to extract, not improve.

OpenAI and Anthropic? racing for capabilities. reasoning, coding, tool use, agentic systems. genuinely impressive work. but self-improvement? behavioral insight? not on the roadmap.

here’s why:

“DO” features are easy to demo. user asks, AI answers. clear input, clear output. great for benchmarks. great for launch videos.

“BE” features are slow. you need to watch someone for weeks to notice patterns. the feedback loop is longer. the wins are quieter. harder to put in a pitch deck.

so we get AI that writes code in seconds but can’t tell you you’ve been losing focus at 14:00 for three weeks straight.

the companies with your data use it against you. the companies building intelligence don’t collect the right data. the gap is wide open.

i’m not imagining this — i’m building it

here’s the stack:

┌──────────────────────────────────────────────────────┐
│                     the stack                        │
├──────────────────────────────────────────────────────┤
│                                                      │
│  ┌─────────────┐                                     │
│  │  Dayflow    │  ← constant screenshots, local     │
│  └──────┬──────┘                                     │
│         ▼                                            │
│  ┌─────────────────┐                                 │
│  │ Gemini Flash    │  ← recognizes what you're doing │
│  │     Lite        │                                 │
│  └──────┬──────────┘                                 │
│         ▼                                            │
│  ┌─────────────┐                                     │
│  │  Clawdbot   │  ← understands context, patterns   │
│  └──────┬──────┘                                     │
│         ▼                                            │
│  ┌─────────────┐                                     │
│  │   self.md   │  ← prediction, nudges, BE layer    │
│  └─────────────┘                                     │
│                                                      │
└──────────────────────────────────────────────────────┘

Dayflow takes constant screenshots and stores them locally. your activity never leaves your machine.

Gemini Flash Lite runs locally, recognizing what you’re doing from screenshots. fast, cheap, private.

Clawdbot — the AI agent I talk to daily — can now see these patterns. it’s the first AI that actually knows how I work, not just what I asked it.

self.md adds the missing layer: prediction. not just “what did you do” but “what will you need.”

the data that powered my 16-day analysis? that’s what feeds this stack. continuously. not a one-time study — an ongoing mirror.

what it looks like in practice

imagine opening your AI on monday morning. instead of blank chat, it says:

→ meeting at 12:30 will cut into your focus window.
  reschedule or protect the slot?

you didn’t ask for this. it noticed.

or after a heavy week:

→ you've worked past 22:00 four times.
  your morning starts have drifted an hour later.
  early close tonight?

or when you’re scattered:

→ you've switched between 6 apps in 15 minutes.
  focus is fragmented. single-app mode for 25 min?

these aren’t hypotheticals. these are predictions my data could power. any AI with access could make these calls.

what would it require?

activity tracking — local. your patterns, on your device. not uploaded for training.

pattern recognition across time. not within one chat. across days, weeks, months. real memory .

context routing. work stuff goes to the coding AI. research goes to the reasoning AI. you don’t explain yourself repeatedly.

portable identity . yours to keep. not locked in ChatGPT. not locked in Claude.

instructions that follow you. your preferences, your style, your rules — everywhere.

a mirror, not a judge. showing you yourself clearly. letting you decide.

the self.md thesis

current AI memory is platform-locked and passive. they store facts. they don’t observe. they don’t predict.

we’re building something different: memory + patterns + prediction. stored locally. owned by you. portable across tools.

the goal isn’t productivity. it’s not doing more, faster.

it’s optimal experience.

Mihaly Csikszentmihalyi called it Flow — those states where challenge and skill align, where you lose track of time, where your best work happens naturally. everyone’s had them. nobody can reliably find them.

self.md wants to help you find them. and protect them.

not just tracking what you did. predicting what you’ll need before you ask. nudging you toward flow. away from drift.

AI that helps you BE, not just DO.

Ray Svitla stay evolving 🐌

Topics: personal-ai self-knowledge behavioral-patterns flow prediction