Praesenti

Name, then which of four skills you want to improve, then a short test: one small, general, easy question, answered using the recording function. That recording is what defines the roadmap for everything after — so the app has heard you before it advises you.

The onboarding got this much weight because iteration 3 proved what happens without it. An exercise whose purpose isn't clear doesn't just confuse people, it destroys their motivation to finish.

Home carries new courses, the training plan and recent progress. Journey carries the level — raised by finishing exercises, which unlocks further ones — with progress broken down by category underneath.

Gamification was researched rather than bolted on: Bartle's four player types, so all of them get something; Csikszentmihalyi's flow, so the challenge tracks ability; and indirect control, because the feeling of choosing freely is what separates a game from a chore.

Monument Valley was the reference for showing continuous progress, but a whole map in a game engine was more than we needed. Plateaus were the answer — stepping stones the fox walks along, ending at mastery.

Speakers rattle through. The exercise hands you a transcript and a limited number of pause bubbles — capped at the number of pauses that actually make sense — and asks you to drag them where you think the speaker should breathe.

The evaluation scores three separate things: the pause placed correctly, the pause placed wrongly, and the pause that should have been there and wasn't. Individual passages of the original speech can be replayed to compare against.

First you imitate: listen to a speaker, then say it back with live feedback on how close you are. Then you compose. A speech you wrote yourself gets a dynamics curve — say the sentence aloud, then drag the handle above each syllable up or down. Holding a handle turns its syllable blue, so the feedback is immediate.

Crescendo and decrescendo are drawn with a finger across the text. The gradient always runs weak to strong from where the finger lands, which means a decrescendo is simply the same gesture right to left.

This screen is where the dark mode was decided. Testing showed the text is far easier to follow when it doesn't move — so it sits still and rotates — and far easier to read against near-black, exactly like a teleprompter.

Stand the tablet in front of you so it can see all of you. The camera reads your posture in real time and maps correction onto you in augmented reality — a prompt with a live correction rather than a verdict afterwards.

Two findings shaped it. Thin lines are almost invisible in AR, so it borrowed the heavier line weight from the dynamics exercise. And because iteration 1 had already established that people won't watch themselves, the camera image is abstracted down to a skeleton — all the information, none of the discomfort.

The design question we never fully settled: whether to show the user what's wrong, or what right would look like.

Gestures reinforce what's said, so the exercise had you imitate a good gesturer to build a repertoire before finding your own style. The interaction was lifted from karaoke: mark in the text where the next gesture belongs, and run a timer on the one after it so you know when it's due.

Testing was blunt about it. Following text that moves is hard; splitting attention across a video, illustrated gestures and a script is harder; and turquoise carrying two meanings at once — progress and gesture position — read as neither.

The findings survived and went straight into the dynamics and posture exercises. The exercise itself did not get finished.