Feature · Coach · BETA
Coach (BETA): record the hold, get a read on your form.
Handstand and planche are the two skills in calisthenics where you cannot self-assess by feel. You can't see your own hip angle. You can't tell whether your shoulders are actually protracted or just tensed. You can film yourself and stare at the frame — but you don't know what "20° torso lean" is supposed to look like, so the video mostly tells you it's a mess.
Coach (BETA) gives you form and technique feedback. Record a 5-15 second hold from the side, and the app returns five readings you can work with.
The five angles
Every coachable movement gets scored across the same five form angles, each compared against a target tuned specifically for the variant you're training:
- Arm lockout — how straight your arms are through the elbow. Full planche wants ~175°; a bent-arm variant tolerates less.
- Torso lean — angle of your body relative to horizontal. Straddle, full, and one-leg planche all run around 85° in profile.
- Shoulder protraction — how far your scapulae are pushed forward. This is the angle most self-taught planche practitioners get wrong.
- Hip angle — for straddle and full variants, how open or folded your hip is. Real straddle geometry runs 160-170°; anything folded shorter reads as a tuck.
- Knee angle — how straight your knees are. Full planche wants ~177°; tuck accepts a bent knee by definition.
Every reading gets a status pill (good / slightly off / far off) and a coaching cue. If the angle is off by 20°, the cue tells you exactly which direction and what to do about it.
Not one ruleset — variant-specific rulesets
The single biggest failure mode in fitness-AI apps is comparing every user against the same platonic form standard. That's how you end up telling an Advanced Tuck practitioner they need to straighten their legs, which is the definition of not doing Advanced Tuck.
Coach's rulesets are indexed by variant. Full Planche and Advanced Tuck Planche use different targets. Wall Handstand and Freestanding Handstand use different targets. The variant hint is captured when you record the session — either automatically from the movement you tapped on mid-workout, or by picking it in the Coach tab — and follows through the whole analysis pipeline.
The current planche rulesets were tuned against a 3-video answer-key corpus of straddle and full planche holds by a real athlete. Readings are a guide, not a verdict. When the angle or lighting makes the pose data weak, the app tells you the readings are rough.
The plain-English coaching note
After the numbers land, a short narrative summarizes what to focus on next. It reads like a coach's cue rather than a spreadsheet: "Torso stayed nice and flat, but hips folded 15° short. Try squeezing your glutes harder to lock the position out."
On eligible iPhones (15 Pro and up), this narrative is generated by Apple's on-device Foundation Models framework — a small language model that runs entirely on the Neural Engine, with no network call. Older devices fall back to a rule-based template that produces the same shape of note from the same reading, so you never get a degraded product for having older hardware.
The prompt itself is strict: single angle, pre-computed delta, direction spelled out in Swift before the model sees it. This is a hard-won rule from a July 2026 spike where letting the model reason across multiple angles caused it to hallucinate arithmetic ("try to lower by 15°" when the numbers were 175, 180, 5). The narrative you get today is on-rails — the model can only phrase what Swift already computed, never invent a number.
Your video stays on your phone
The full pipeline runs locally:
- Pose detection — Apple's Vision framework, on-device.
- Stable-window detection + angle math — pure Swift, on-device.
- Ruleset comparison — bundled JSON, on-device.
- AI narrative — Apple's Foundation Models framework, on-device on eligible iPhones. Rule-based template fallback for older devices.
Coach saves the numeric readings to your account (so trend charts in the Coach tab can render your progress over time), but your form video and the body keypoints the analysis reads from it never upload. We only ever receive the angle measurements and the hold time.
Two ways to reach Coach
Coach tab (5th tab). Pick a movement, record fresh or import from Photos. Trend charts show how each angle has evolved across sessions.
Mid-workout "Coach this hold" button. On any coachable movement's set row during a session, tap the button to jump into Coach with the movement + variant already selected. Capture, get the read, save, and return to the workout without leaving the flow.
What's coming
Right now Coach covers handstand and planche variants. Future expansion:
- More skill families — front lever and back lever are the natural next candidates. Same 5-angle pipeline, tuned rulesets per variant.
- Compare to a pro — side-by-side split-view with a reference athlete's clean hold, playhead-synced. Eventually with the pose overlay rendered on both.
- Reference videos per step — a 3-5s loop of what a clean version of each movement looks like, so you can preview before you record yours.
These are gated on adoption, not code. If Coach becomes something you use every week, the case for building the next tier is easier to justify.
Free, on your phone
Coach is included in the free app. Download, pick a movement, film 10 seconds from the side, and see what your hold actually looks like.
Frequently asked questions
How does Coach analyze my handstand or planche?
You record a short clip from the side (5-15 seconds is ideal). Coach uses Apple's Vision pose-estimation to track joint positions across every frame, detects the stable window where you're actually holding the position (as opposed to kicking up or falling out), and computes five form angles: arm lockout, torso lean, shoulder protraction, hip angle, and knee angle. Each angle is compared against a tuned ruleset for the specific variant (tuck vs advanced tuck vs straddle vs full planche, wall vs freestanding handstand). You get a numeric reading, a status pill (good / slightly off / far off), and a coaching cue for each one.
Does my video leave my phone?
No. The whole analysis pipeline — pose detection, angle math, ruleset comparison, and the AI-generated coaching note — runs on-device. Vision runs locally; the AI narrative uses Apple's Foundation Models framework on eligible iPhones (15 Pro and up), which processes your prompt data on the Neural Engine without any network call. Your form video and pose data stay on your phone; the angle readings and hold time are saved to your account so you can track progress. Coach is in BETA and gives form and technique feedback: when the angle or lighting makes the pose data weak, it tells you the readings are rough.
Which movements does Coach cover?
Handstand variants (wall handstand, freestanding, L-shape, hollow-body) and planche variants (straddle, advanced tuck, tuck, full). Each variant has its own tuned angle targets — a Full Planche and Advanced Tuck aren't graded against the same standard because they're not the same position. You reach Coach two ways: the dedicated Coach tab, or a 'Coach this hold' button that appears mid-workout on the set row of any coachable movement so you can capture the analysis in the flow of a session.
What's the plain-English coaching note?
After analysis lands, Coach produces a short narrative summarizing what the numbers mean and what to try next. It reads like a coaching cue, not a report: 'Torso stayed nice and flat, but hips folded 15° short. Try squeezing your glutes harder to lock the position out.' On eligible iPhones this narrative is generated by Apple's on-device Foundation Models with a strict single-angle prompt (never multi-angle — the model gets one focused signal, not a scatter). Older devices get a rule-based template that produces the same shape of note from the same reading.
How much does Coach cost?
Nothing. The whole app is free — every program, every calibration, every Coach analysis. No subscription, no paywall, no in-app purchase gating any of it.
