01 · Product Vision

1. The problem: great coaching does not scale

A great personal coach is not an information product. The information — sets, reps, macros — has been free on the internet for twenty years. What clients actually pay for is a relationship: someone who knows their history, notices when they slip, adjusts the plan before Monday goes wrong, and whom they do not want to disappoint.

That relationship is expensive to produce. Nata's 1:1 remote coaching today (numbers are her real workload, lightly rounded):

At ~110 minutes per client per week, a full-time coach saturates at 12–15 clients. Nata is at 14 and turning people away. Raising prices filters people out; hiring junior coaches dilutes exactly the thing clients came for — her.

The industry's answer — generic fitness apps — solves the wrong problem. Apps scale content, not relationship:

• Nobody is watching. The app does not care if you skip Tuesday. Typical consumer fitness apps lose the majority of users within the first month (industry-typical 30-day retention is ~25–30%; assumption, directionally well-established).
• The user does the work. Logging food, choosing weights, deciding what to do on a bad-sleep day — the app pushes labor onto the person least equipped to do it consistently.
• No memory that compounds. The app on day 300 knows barely more about you than on day 3.

So the market splits into two bad options: a $15/month app that ignores you, or a $300/month human who cannot take you. NataCoach is built for the gap between them.

2. Who Nata is, and what only she brings

Nata is a real strength and nutrition coach (9 years coaching, ~200 clients over her career; current specialties match our two pilot personas: postpartum return-to-strength and fat loss for desk-bound professionals — see Marta & Denys). Her clients stay for years, and when asked why, they do not say "the programs." They say: "She notices."

What only Nata brings, and what the system must never fake:

1. Programming judgment. Turning "8 months postpartum, wants a pull-up, sleeps badly" into a periodized 12-week program is expert work. The AI executes and adapts programs; it does not author them.
2. Judgment at the edges. Pain that might be an injury, a plateau that might be underfeeding, a client quietly spiraling — pattern recognition from a decade of humans, not a dataset.
3. Presence. A real voice note from Nata after your first full pull-up negative is worth fifty perfectly worded bot messages. Her authentic presence is the moat, and we spend it deliberately (see Coach-in-the-loop).
4. Permission to be trusted. People accept "cut today's volume 20%" from someone they believe has earned the right to say it.

3. The product thesis

AI handles the every-day surface area. Nata supplies programs, judgment, and presence.

Roughly 80% of Nata's 110 weekly minutes per client is surface area: reading, acknowledging, computing, proposing, reminding, summarizing. Almost none of it requires her judgment — but all of it requires someone, every day, or the relationship decays. That surface area is exactly what an LLM with a persistent per-client memory (the Wiki Brain) does well.

Target end state: Nata's time drops from ~110 to ≤ 30 min/client/week in the pilot (and ≤ 10 at scale), while the client experiences more daily contact than a 1:1 human coach could ever give.

4. The seven locked decisions — deliberate improvements over the original ask

The original request (brief §2) was directionally right. Where we changed it, we changed it on purpose:

5. Product principles

1. The user does less. The prime directive. Every flow is scored by taps required and seconds spent. If a feature adds user labor, it is wrong until proven otherwise. Target: an engaged day costs the client under 3 minutes of interaction.
2. Chat-first. Plain Telegram chat — photos, buttons, voice — is the product. The Mini App is for moments a chat genuinely cannot serve (live workout runner, charts, Console). If it can be a message with two buttons, it is a message with two buttons.
3. Propose-confirm. The system computes the default; the human taps ✓ or adjusts. This applies to clients (weights, macros, weigh-ins) and to Nata (drafted feedback, suggested program tweaks). Nobody starts from a blank input field.
4. Coach-in-the-loop. Autonomy is a dial, not a switch, set per message type and moved only when edit-rate data says the AI has earned it. Safety-critical messages never leave escalate.
5. Compounding memory. Every interaction makes the next one better, because everything distills into the Wiki Brain. Month 6 must feel noticeably more personal than week 1 — that is the retention engine and the defensibility.
6. Safety first. When in doubt: pause the plan, escalate to Nata, say less. Estimates are labeled as estimates. The bot never diagnoses.

6. The core value loop

The loop everything else in these docs serves:

flowchart LR
 A["Data in<br/>one-tap confirms, meal photos,<br/>wearable sync, chat"] --> B["Wiki Brain<br/>distilled, current model<br/>of this client"]
 B --> C["Better coaching<br/>recovery-aware, specific,<br/>in Nata's voice"]
 C --> D["More trust<br/>advice lands, effort feels seen,<br/>confirming stays effortless"]
 D --> A
 N["Nata<br/>programs, edits, voice notes"] --> B
 N --> C

Generic apps never close this loop: data goes in, nothing personal comes back, logging stops by week 3. NataCoach closes it daily — and Nata's edits inject ground truth the loop could never learn on its own.

What the loop feels like on a random Tuesday (Marta, week 5 — full transcripts in User Experience):

07:10  @NataCoachBot: Morning, Marta ☀️ Whoop says recovery 44% — rough
       night (2 wake-ups, 5h 40m sleep). Today was Pull B. I'd keep the
       pull-up negatives but drop rows from 4 sets to 3 and skip the
       finisher. Sound right?
       [ ✓ Do that ]  [ Keep full plan ]  [ Move to tomorrow ]

07:11  Marta: ✓ Do that

07:11  @NataCoachBot: Done — 35 min instead of 50. See you at the bar 💪

Two taps. The plan adapted, the event was logged, the Wiki Brain got a little sharper about how Marta's bad nights look, and Nata spent zero minutes.

7. Positioning

The wedge: coach-grade personal attention at roughly one-third the price of 1:1 coaching, at 10× a coach's capacity — a segment neither incumbent can serve.

8. Pilot success metrics (10 users, 8 weeks)

All metrics are computed deterministically from events (locked decision 6), visible live in the Coach Console. Targets are honest bars, not vanity:

Secondary signals we watch but do not gate on: median user taps/day (expect 6–10), Morning Brief response rate, voice-note reactions, unprompted messages to the bot (a trust proxy — we want these to grow).

9. Non-goals (for now)

Per the locked technical decisions (brief §6): no custom pose-estimation models (vision LLM + checklists first — see Form Check), no barcode scanning, no native iOS/Android apps, no payments until Phase 3, no multi-coach marketplace. NataCoach scales Nata; generalizing to "any coach" is a question for after the pilot proves the loop.

Next: how this vision becomes daily experience in 02 · User Experience, and how it gets built in stages in 11 · Roadmap.