KEELa body of evidence Working document · 22 Aug 2026
Personal wellness operating system

One day in.
Everything out.

Talk your day into it once. Get back the evidence of a life, for as long as you keep going.
14Straight days
at the gym
234Days of 10K steps
unbroken since Jan
40kgLost since
Jan 2024
0Apps that hold
any of it
01

Why this exists

The habit came first. The app is just the part that remembers it.

This did not start as a product idea. It started on 22 August 2026, in the bath, after an upper-body session and a 10K walk, on the fourteenth consecutive gym day. The thing worth noting was not the fourteen. It was that for the first time going felt good rather than forced.

That is a new state, and new states are worth instrumenting before they fade. There is no job at the moment and the runway is tight, and this still reads as the more important thing happening right now.

So Keel is not being built to start a habit. The habit already exists. The data is already being produced every single day. It is just landing in four different apps and in nobody's memory, which means none of it compounds.

02

One way in.
Many ways out.

Every wellness app dies in the same place: the form. Twelve fields, four dropdowns and a date picker, on a phone, sweaty, at 11pm. Nobody sustains that.

So Keel inverts it. Input is conversation. Say the day the way you would tell a friend, out loud if that is easier, and the app does the structuring. Output is structure. Tabs, charts, streaks, summaries, every one of them a read-only view built from the same log.

The test for every feature that ever gets added: does it give you another thing to fill in, or another thing to see? Filling in needs a very good reason. Seeing does not.

03

How it actually works

This is the whole product. One paragraph of ordinary speech at the end of a day, resolved into records. Press the button.

Voice note · 22 Aug · 21:48

I woke up today at 6am. I went to the gym, did upper. Before going to the gym I took creatine and a banana, then came back home, did 10K steps, then took a warm shower, and I had a protein before the walk.

Extracted

    Nothing logged yet.
    Six records are waiting in that paragraph.

    Every extracted number keeps the sentence it came from. Tap any record in the real app and the original words light up, exactly like above. Nothing is saved until it has been confirmed on screen. Silent wrong parsing is the one thing that would end the trust, and trust is the entire product.

    04

    What goes in

    Eleven data types. Each one has a home, a history and a chart.

    01
    Body measurementsHeight, weight, waist, chest, arms, thighs, neck. The full set monthly, prompted, not remembered.
    02
    Daily weightOne number, five seconds. Shown as a seven-day trend, never as the raw daily figure.
    03
    WorkoutsSets, reps, weight, volume, PRs, muscle group balance. Imported from Hevy or spoken.
    04
    StepsThe oldest streak in the system and the most prominent counter in the app.
    05
    WaterOne tap per glass. The single place where a button beats a sentence.
    06
    FoodFree text or a photo. Protein is the number that matters. Calories are estimates and are labelled as such.
    07
    PhotosAny number a day. Progress shots frame against a ghost of the last one so the comparison actually works.
    08
    Journal and moodFree writing, any length. Read by the summaries so a bad week gets answered like a bad week.
    09
    SupplementsCreatine, multivitamin, omega 3, D3. Recurring, so it pre-fills instead of being typed.
    10
    ScansDEXA and blood work as milestones, not rows. Upload the PDF, compare scan against scan.
    11
    SleepNot in the original list. The strongest predictor of everything else in it.
    05

    The screens

    TodayFour quick actions, the input box, today as it accumulates, both streaks above the fold.
    DiaryEvery day gets a page. A calendar for scanning a month, a feed for reading it back.
    ExercisesSession history, per-lift history, PRs, volume over time.
    FoodMeals by day, protein against target, the photo grid.
    BodyMeasurements, weight trend, progress photos side by side, scans compared.
    InsightsNightly, weekly, monthly and yearly summaries.
    AskA question box over your own history, with the days cited.
    SettingsTargets, tone, privacy, spend cap, export everything.
    06

    What it says back

    A summary every night, every week, every month, and every year from the day it starts. What is going right, what is going wrong, in plain words.

    Three rules make it worth reading. It has to be specific, so "protein landed at 52g, third day under 80" instead of "try to eat more protein". It gets a maximum of two things to fix, because a list of nine is a list of zero. And it reads the journal first, so a day that was written up as a low day gets a different answer than a day that was not.

    Tone is a setting, from gentle to blunt, changeable mid-week. Some weeks need you missed two sessions, go tomorrow. Some weeks need you held the streak through a job hunt, that counts. Same data. The app should not have to guess which one is wanted.

    Correlations, sleep against lift performance, protein against mood, stay switched off until there are sixty days of data. Correlations on three weeks are noise wearing the costume of insight, and one confident wrong claim early would poison every true one after it.

    07

    The hard part, told straight

    Pulling data out of fitness trackers is the part most likely to disappoint, so here it is without optimism. Every line below needs testing before anything is built on it.

    SourceWeb app todayNeeds a native appFallback
    HevyCSV export, imported by handNoSpeak it
    Apple HealthExport zip only, no server API existsYes, for live syncSpeak it
    Health ConnectNo. On-device onlyYesSpeak it
    Weight scaleManualDepends on the scaleType it
    DEXAUpload the PDFNoType the numbers

    Apple and Google both keep health data on the device on purpose. No website can read it, at any price, which is the real reason a native app eventually exists. Notice that the fallback column is identical on every row. That is not a backup plan, it is the plan, and it is exactly why the talking has to be excellent.

    08

    Build order

    Each phase has to be usable on its own, and nothing starts until the one before it has been used for real for a week.

    NOW
    Phase 0 · The baseline Book the DEXA. Take every measurement. Photos from four angles. Write the starting numbers down. This happens whether or not any code ever exists, and it is the only part that cannot be recovered later.
    01
    The logAuth, the input box, parsing, the confirmation screen, Today and Diary. Nothing else. This is the whole idea in embryo and it either feels good or the idea is wrong.
    02
    Voice and photosSpeaking instead of typing, attachments, the photo pipeline.
    03
    The viewsExercises, Food and Body tabs. Charts, trend line, streaks.
    04
    IntelligenceThe nightly summary, then weekly. Ask your own data.
    05
    ImportsHevy CSV, then health exports.
    06
    The long viewMonthly and yearly summaries, and the weekly share card.
    07
    NativeOnly after a month of real daily use, and only because Health Connect leaves no other route.
    09

    Open questions

    Parked, not blocking. Answer them whenever.

    1. Which DEXA clinic, what does it cost, when is it booked? The baseline waits on it.
    2. Is Hevy Pro worth paying for automatic sync, or is a manual CSV export fine?
    3. One user forever, or others eventually? It changes auth, schema and legal, and it is far cheaper to decide now than to retrofit.
    4. Browser speech-to-text, or a paid transcription API? Needs a real test at real speed.
    5. How blunt should the coach be by default, before the setting gets touched?
    6. Progress photos daily, weekly, or monthly?
    7. What is the actual goal? A weight, a lift, a body fat percentage, or just continuing? The summaries need something to measure against, and "keep going" is a legitimate answer if it is the true one.
    8. Does the streak survive a rest day? Define what breaks it before one breaks by accident.
    9. Does the app itself appear in the public content, or only the results?
    10. Run a proper trademark search on the name before anything commercial.