Two minutes a day: the cognitive math behind a 14-day reflection arc

A daily mental-health check-in app has about ninety seconds of patient attention before it loses the day. Maybe a hundred and twenty on a good day. After that, the patient closes the app, the entry does not happen, and the streak (if there was a streak) is broken, and tomorrow becomes harder, not easier.

The 14-day arc in Nyra was designed against that budget. Two minutes a day, ceiling. Eighty percent of days closer to thirty seconds. No streak count, no reminder cascade, no in-app reward animation when you finish. This post explains why every one of those choices was load-bearing, and what the math is behind the cadence.

The attention budget for a wellness check-in

Patients in outpatient behavioral-health care are not bored. They have jobs, family, school, a commute, a phone full of other apps. The window for a daily reflection competes with every other thing that wants ninety seconds of their attention.

The empirical work on app retention is unambiguous about what kills retention: friction in the first interaction, escalating ask in subsequent ones, and the loss-frame mechanics (streaks, milestones, missed-day notifications) that turn a self-care behavior into a debt.

So the design constraint we wrote down on day one was: the patient must be able to close the app inside ninety seconds with the day's entry complete. Anything that makes the ninety-second close-out impossible was on the chopping block. A lot of features did not survive.

What survived is a single prompt. "One moment from today." The patient writes (or speaks) two sentences, two paragraphs, whatever is honest, and closes the app. Done.

The signal extraction happens server-side, asynchronously. The patient does not wait for it, does not see it, does not tune it. They wrote about today; the rest is the product's job.

What we did not ship

The list of features we did not ship is the part of the design that matters most.

No mood slider. A 1-to-10 mood slider feels like a low-cost ask. It is not. It turns a reflection into a data-entry task, and the patient who has to "rate" their day before describing it will rate it before describing it for the rest of their relationship with the product. The rating bias affects the reflection that comes after. We let the patient write first and rate nothing.

No category tags. No "is this about work, sleep, family, or other." Categories sound helpful and are quietly extractive. They train the patient to pre-classify, which is the exact opposite of what reflection is for. The signal extraction reads the open text and finds the categories; the patient does not see them.

No streak count. The hardest decision and the right one. Streaks work on a calorie-tracker. They actively harm a mental-health check-in. A patient who has had a hard week and missed three days is not motivated by the loss of a 12-day streak; they are motivated by being able to come back without the product reproaching them. We do not show a streak number anywhere in the app. Missed days do not get a red mark. The 14-day arc has a finish line; the streak does not.

No missed-day notification. "We missed you yesterday." This sentence appears in most consumer health apps. It is also the single most reported reason patients delete those apps. We send no missed-day notification. The only notification the patient gets is the daily prompt, sent at a time they configure once during onboarding, and even that can be turned off.

No in-app score chart. The patient does not see their PHQ-9 trajectory plotted against population norms. The literature on self-administered instruments is clear that score self-interpretation, in the absence of a clinician, can produce worse outcomes than no measurement at all. The chart is for the clinician's review; the patient sees a confirmation that their response was recorded.

Why fourteen days

The 14-day arc lines up with the visit cadence in outpatient psychiatric and behavioral-health care. It is short enough to feel finite. It is long enough to capture a real pattern. It gives PHQ-9 and GAD-7 three meaningful checkpoints (day 1, day 7, day 14) instead of one. We wrote a longer post on the 14-day case that covers the clinical rationale; here we focus on what the patient experiences.

The patient gets, on day one, the explicit commitment: reflect for fourteen days, then your clinician walks into your next visit prepared. The arc has a defined end. The patient is not signing up for "indefinite mental-health tracking"; they are signing up for two weeks. That framing changes adherence.

At the end of the arc, the next arc starts at the next visit, on the same fourteen-day cadence. There is never a moment where the patient is "expected to engage forever." Engagement is per-arc, and each arc has a payoff at the next visit.

How voice changes the cognitive load

The single largest cognitive-load reduction we shipped was voice. The patient holds the record button, says two sentences, releases. Done. The interaction is faster than typing for most patients, and for patients with dyslexia, motor impairments, or English-as-a-second-language, the gap is much larger.

Three constraints on the voice surface:

Transcription is on-device, when the device supports it. Modern iOS and Android have local speech-to-text APIs that produce clinical-quality transcripts without ever leaving the phone. The transcript is what the server sees; the audio file is discarded.

On devices that do not support local transcription, we fall back to a server-side model. In that case, the audio is transmitted over TLS, transcribed in a stateless function, and the audio is discarded after the transcript is produced. We do not store the audio. We do not pass the audio to a third-party model that retains it. The transcript is the artifact.

The patient can edit the transcript before submitting. Voice-to-text is not perfect. The edit-after-record flow lets the patient correct anything that came out wrong, which keeps the patient in control of what their clinician will read.

The net for the patient: thirty seconds of voice instead of two minutes of typing, with the same end state.

What the math actually says

A few back-of-the-envelope numbers, because the design pretends to be quantitative.

The median text reflection in our pilot is 38 words. At a moderate typing speed of 30 words per minute on a phone, that is 76 seconds of typing, plus ~10 seconds to launch the app, ~5 seconds to read the prompt, and ~5 seconds to confirm submission. Total: 96 seconds. Inside the ninety-second budget on a good day, just over on others.

The same reflection in voice is roughly 18 seconds of speech, ~4 seconds of transcription processing, plus the same overhead. Total: 42 seconds. Comfortably inside the budget.

On scale days (1, 7, 14), the additional ask is PHQ-9 (9 items) or PHQ-9 + GAD-7 (16 items) on day 14. At ~3 seconds per item to read and respond, plus ~2 seconds between items, that is ~27 seconds for PHQ-9 and ~48 seconds for the day-14 pair. The scale-day ceiling becomes 96 + 27 = 123 seconds for typed reflection on PHQ-9-only days; 42 + 27 = 69 seconds for voice reflection on the same days.

We tell patients "two minutes a day." The math says ninety seconds median, two minutes on the worst combination (typing + day 14 scales). The promise undersells slightly, which is the side of the asymmetry we want.

The tradeoff we made on signal

The cost of the ninety-second budget is that we extract less signal per day than a "full" mood-tracking app would. We do not ask about sleep specifically. We do not ask about appetite. We do not ask about social contact, exercise, substance use, or any of the dimensions a full daily questionnaire would cover.

The bet is that fourteen days of honest reflection extracts more reliable signal than fourteen days of fatigued data entry on a questionnaire. The reflection captures the salient thing of the day (the patient picks); the classifier reads sleep, appetite, mood, energy, anxiety, anhedonia, suicidal ideation out of the text. The validated instruments on days 1, 7, 14 ground-truth the classifier against scales the literature already trusts.

That is the architectural reason for the design: the open prompt is the upper bound on reliable reporting, the closed instrument is the calibration. The clinician sees both.

What changes for clinicians

The patient experience designed for ninety-second budgets has downstream effects on what the clinician sees. The evidence map is built from reflections that exist because the patient was actually willing to write them, not from a questionnaire that was fatigue-completed. A 14-day arc with twelve days of honest two-sentence reflections is more clinically useful than the same arc with four days of complete questionnaires and ten days of skipped entries.

The clinician's pre-visit summary points at the moments. Not at "PHQ-9 trend down 3 points," though that is in the chart. At the moments: "Sleep disruption signal, days 3-6, source rfl_4a91." The summary is grounded in things the patient wrote, in two-sentence chunks, at ninety seconds a chunk.

Where to go next

If you are a patient curious what this actually feels like on the phone, the app is invite-only through your clinic right now. Ask your clinician whether they are taking part, and read more about the patient experience.

If you are a clinician thinking about how this would land in your practice, book a thirty-minute walkthrough. The walkthrough includes the patient flow, the clinician evidence-map surface, and the audit posture.

For the wider context, see the case for 14-day mood check-ins. For the safety posture that runs underneath the daily reflection, see the item-9 escalation design.