Morpheus — Product Spec

A diet companion in iMessage that remembers what you eat, tells you what to order, and watches you change. The ideal product.

2026-06-17 · updated 2026-06-18 · companion to DESIGN.md · v0 = single-user dogfood

Thesis

Dieting fails on accountability and what goes unspoken, not on tracking. Tracking is a solved commodity that leaves you alone with a passive log. Morpheus is the relationship: it remembers what you eat, tells you what to order when you ask, and watches how you look week-to-week — and that alone bends the curve.

• Attention beats data.Perfect macros you ignore change nothing. A friend who notices you've been softer this week — and says so — does.
• The body is the ground truth.Most people don't own a scale and don't want one. A photo trend over weeks is honest signal — often more honest than a number.
• Two signals are enough.How you look + what you ate, tracked over time. You don't have to do much; Morpheus needs nothing else.

Onboarding built

A short, casual Q&A — one question at a time — then it wires up tracking and confirms it can see you. No forms.

• Learns the basics.Goal (lose/maintain/gain + target), starting weight, usual meal times, and a weigh-in time — saved to the profile as answers land.
• Links your location.Asks you to Share My Location in Find My; the moment you do, it finds your share among everyone sharing with the account, links it to you, and learns your city.
• Confirms it's real."ok yeah, i can see you're in fremont. good, we're set" — the eval. Then the daily ritual begins.

What Coach does built

One job, done well — and nothing else.

• Talks like a person.gen-z tough-love, short texts, multi-bubble, human typing cadence. The relationship is the product.
• Logs food.Mention a meal or snap a plate → rough macros, portion call-out, filed against the day and the restaurant.
• Reads the mirror.Progress photo → compares to your recent check-ins → verdict + observation on a timeline.
• Tells you what to eat.Suggests meals you like that fit the day's remaining calories + protein, craving-aware.
• Catches you in the wild.Continuous Find My location → dwell + venue detection → proactive callout.
• Stays in its lane.Strictly a diet/body coach. Ask it to do your math homework and it bounces you in character ("not really what i'm for, what'd you eat today") and never breaks role.

The daily ritual

The whole product is one small, repeatable habit — it survives because it takes ten seconds and someone's waiting for it.

• Morning mirror pic.A shirtless / progress photo, daily, same time and light. Morpheus stores it and compares it to the last ones.
• Photo of what you eat.Snap the plate. Morpheus gets a rough calorie baseline, flags the portion, files it against the restaurant and the day.

The mirror — visual progress hero · built

Send Morpheus a photo of yourself; it judges this one against last time and tells you whether you progressed. The scale becomes optional — the camera is the instrument.

• Compares, doesn't grade.Reasons about the delta vs. your last few check-ins (leaner / softer / flat / more definition). No creepy score.
• Writes notes it keeps.Each check-in produces observations ("waist tighter, posture better") stored on a timeline it references next week.
• Cross-references the food.If the photos stall and the log shows portions creeping up, it connects the dots out loud.
• Weight is optional.Got a scale? One more signal. No scale? The body trend carries the whole thing.
• Reasons on the trend, not consecutive frames.Daily noise (pump, light, time of day) is smoothed over multiple days so it doesn't over-react.

How it works: the inbound photo is classified (body / food / other); a body shot is sent to the model alongside your last couple of check-in photos so it reads the delta, then writes the verdict to the timeline. Photos stay on the self-hosted Mac mini as local file refs (HEIC→JPEG via sips) — sent to the model for the read, never to a third-party app.

Ask it what to eat main value prop · built

The core loop: Morpheus remembers what you actually like, so when you're hungry it doesn't hand you a meal plan — it suggests something you'd genuinely eat that still fits your goal.

• "what should I eat?"Answers from your own tastes and history, not a generic database — meals you've liked before, tilted toward the day's remaining calories and protein.
• Craving-aware."I'm feeling Korean" → a specific dish that scratches it and fits the goal: "bibimbap, lean beef, easy on the sauce — high protein, ~550."
• Proactive, not just on-demand.It asks at meal times ("what are we eating?") and steps in the moment you say you're hungry — before the bad decision, not after.
• Learns from what you pick.Every accepted or rejected suggestion sharpens the next — it converges on the short list of goal-friendly meals you'll actually repeat.

Location callouts & restaurant memory built · Find My

Morpheus watches where you are and texts first when you've parked somewhere you shouldn't — and because it has learned your spots, it names your usual order. No app, no Apple Shortcut: it reads the location you share over Find My, continuously.

• Continuous, straight from Find My.You Share My Location with Morpheus once; it reads your live coordinate on the Mac mini and tracks you — no geofence Shortcut to install.
• Dwell, not drive-by.It only calls out after you've STAYED a while, so passing a McDonald's doesn't trigger it — parking at one does.
• Knows the spot even if you never told it.It reverse-geocodes where you are; a fast-food brand or category → "why are we at mcdonald's" — no pre-saved danger list needed.
• Names your usual at spots it knows.For a learned restaurant it cites the usual order and offers the lighter swap; a one-word "yes" logs it from memory — no photo, no typing.

Mechanism: the phone shares location over Find My; findmylocateagent's encrypted SQLite is decrypted on the Mac mini (Apple's AES-256 page codec, 32-byte key grabbed once via lldb), the live coordinate read from the WAL, dwell-checked, then matched to a saved restaurant or reverse-geocoded (OpenStreetMap). The old Apple-Shortcut geofence + HTTP endpoint were removed.

Nutrition lookup built

You never count. Morpheus figures out the numbers — from the web, the photo, or what it already knows.

Memory model built

The store grew from a thin profile keyed by handle into a lightweight graph — records that reference each other by id (no database until it earns one). It now holds restaurants, usual orders, dishes, the daily meal log, and the body-photo timeline, plus the linked Find My sharer id per user.

user {
  profile:  { goal, restrictions, baseline, tone }
  patterns: [ "late-night burritos on weekends", ... ]
  restaurants: { "chipotle-mission": {
    geo:{lat,lng}, dangerLevel:"high", visits:14,
    usualOrders:[ { dish:"carnitas-bowl", freq:11, calEst:900 } ] } }
  dishes: { "carnitas-bowl": {
    calEst:900, macros:{p,c,f}, source:"search→photos", confidence:0.8 } }
  meals: { "2026-06-17": [ { time, slot, dish, calEst, photo,
    portion:"large", restaurant:"chipotle-mission" } ] }
  checkins: [ { date, photo, observations:["waist tighter"],
    verdict:"progress", comparedTo:"2026-06-03" } ]
  weight?: { "2026-06-17": 174 }   // optional
}

Photos stored as local file refs on the Mac mini, not blobs in the JSON.

A day

Architecture

Forked from a self-hosted iMessage agent stack. Transport and the agent loop were already proven; the net-new work — vision, the memory graph, nutrition lookup, and continuous location — is now built.

The inbound photo router

Build order

Ship the relationship first — if the conversation isn't good, no camera or location trick saves it. All five v0 steps are now in.

Open questions

• Photo classification.resolved Auto — a vision classifier decides body / food / other and routes the photo.
• Mirror cadence.resolved Daily, folded into the weigh-in nudge; reasons on the multi-day trend to avoid over-grading noise.
• Danger zones.resolved No hardcoded list — live reverse-geocode flags fast-food by brand/category, gated behind dwell.
• Find My sharer mapping.open Today the user's findMyId is picked at onboarding (freshest fix) and pinned; auto-maps by handle once Apple's friends table syncs.
• Confirm vs. react on the mirror.open Should daily check-ins always get a real reaction, or a lighter ack on flat days?