Sleep Cycle

Promptfree, for everyone, and the only version there is

Copy the prompt for Claude CodeCopy the prompt for CodexCopy the prompt for Cursor

Copy

Build me a sleep tracker that replaces Sleep Cycle — and read the honest part first.

**Sleep phase detection from sound and movement is an approximation**, not a measurement. The paid product's version is tuned against a great deal of data; yours will be cruder, and neither is a sleep study. Build it for the one thing it genuinely does well — waking you at a lighter moment within a window, which most people notice — and be careful not to present a guess as a diagnosis.

STACK
- A native application, because it must run all night with the screen off and access the microphone or the motion sensors. A web application cannot do this
- SQLite on device
- All processing local. Audio never leaves the phone, and that is not negotiable

THE ETHICAL FLOOR
- This records the microphone in a bedroom all night. Process audio in memory, extract features, and discard the raw audio immediately unless the user has explicitly asked to keep a snore recording — and then only that clip
- Never upload audio. Never send it to a service for analysis. Say so plainly at setup and in the README
- Do not present any of this as medical information. Sleep apnoea, insomnia and the rest are diagnosed by clinicians, and an application that implies otherwise causes real harm. If the data suggests something unusual, the only appropriate message is a suggestion to speak to a doctor

THE DATA MODEL
- sessions: id, started_at, ended_at, timezone, alarm_window_start, alarm_fired_at, wake_method, quality_rating, notes, source
- samples: id, session_id, at, motion_magnitude, sound_level, feature_json — sampled every thirty seconds or so, not continuously
- phases: id, session_id, from_at, to_at, phase, confidence — derived, and recomputable from samples
- events: id, session_id, at, kind, clip_path — a snore, a cough, a noise
- factors: session_id, kind, value — caffeine, exercise, alcohol, screen time, whatever the user chooses to log
- Everything derived is recomputable from samples, so improving the algorithm improves the history rather than orphaning it

DETECTION, HONESTLY
- Movement from the accelerometer, or sound level from the microphone, sampled at intervals. Both are proxies for restlessness
- Light and deep sleep separate reasonably well from movement patterns; REM does not separate reliably from either signal, and claiming otherwise is where these applications overreach. Report what you can defend and label confidence
- Keep the feature extraction and the classification separate, so the model can be replaced and the history recomputed
- Show the confidence on the chart. A hypnogram drawn with a confident line from a noisy signal is a picture that misleads

THE ALARM, WHICH IS THE FEATURE THAT WORKS
- A window rather than a time: wake me between 06:30 and 07:00, at the lightest moment detected
- If nothing lighter is found, fire at the end of the window. Never late, ever — that is the contract
- Gentle sound rising in volume over half a minute
- It must fire even if the application has been killed, the phone has been rebooted, or the battery is low. Schedule it with the platform's own alarm facility rather than relying on a running process, and test all three cases. An alarm that does not go off is the only unforgivable bug here
- A backup alarm at the end of the window, independent of everything else

WHAT IT SHOWS
- Time in bed, estimated time asleep, time to fall asleep, and wake-ups. Those four are defensible
- A chart per night, and trends over weeks — which are far more meaningful than any single night
- Correlations with logged factors, presented as correlations and never as causes, with the sample size shown
- A consistency measure: going to bed and waking at similar times is the thing most strongly associated with feeling rested, and it is the one number worth nudging

BATTERY AND TRUST
- All-night recording must cost a small percentage of the battery. Sample sparsely, avoid keeping the screen or the processor awake, and measure it
- A phone that is flat in the morning is an application that gets deleted the same day

ESCAPE
- Export every session and every sample as CSV in two taps
- Write to the platform's health store so sleep appears alongside everything else, and read back what other applications recorded
- No analytics in the application. Not one event

WHAT MATTERS MOST
The alarm firing reliably, and battery cost. Get those two right and the tracking can be imperfect; get either wrong and nothing else matters.