The Maestro bet: any skill, on demand, no wedge

Duolingo started with Spanish. Yousician started with guitar. Tonal started with strength training. The logic was correct for them. Content was expensive to make: domain experts had to record every lesson, every variation, every drill. With expensive content, the right strategy was to get exceptional in one category, build a moat in that one vertical, and only expand once the unit economics were proven. The 21st skill cost the same as the 5th, which is to say a lot.

That ground truth changed in 2025.

A multimodal LLM with vision and reasoning can produce a 14-lesson curriculum for any physical skill in under sixty seconds, given a skill name and a small reference corpus. MediaPipe gives me hand and pose tracking on-device in the browser for free. ElevenLabs gives me a calm coach voice for cents per minute. The cost of supporting a new skill collapsed from a quarter of content engineering to an API call.

When the cost of content production approaches zero, the wedge logic stops being correct.

What the new strategy looks like

Maestro's home page does not show a list of skills. It shows a microphone. The user says what they want to learn, or holds an object up to the camera. The agent identifies the object, generates the curriculum, and starts coaching live. Mid-session, the user can say "switch to juggling" and have a juggling coach in two seconds.

The dynamism is the product. The wow moment is not a single skill done well. The wow moment is the system handling whatever the user puts in front of it.

The architectural consequence: there is no "lesson screen" pre-designed for a specific skill. There is a small set of view primitives (camera-view, reference-clip, ghost-overlay, side-by-side, top-down, annotation, plan-card) that the agent composes per turn for whatever skill is currently active. The same primitives serve harmonica practice, juggling drills, sourdough proofing, and knife-skill knife-work. Adding the 50th skill costs nothing in design or engineering. The skill is just another input to the curriculum generator.

The business consequence: the unit economics change. In a wedge-bound app, you spend $X per skill to produce content, recover it via subscriptions in that skill's vertical. In a wedge-free app, you spend nearly nothing per skill, and your revenue scales with breadth instead of depth. The per-user ARPU lands similar to Duolingo Plus. The total addressable market is every craft a human has ever learned instead of one vertical.

Capture mode: the engine, run on observation

Beyond the user requesting a skill, Maestro can learn one by watching. Capture mode: point the camera at anyone doing a specific skill for thirty minutes. The agent watches, segments the actions, extracts the moves, generates a 14-lesson curriculum keyed to that specific person's technique.

This is not a separate feature. It is the same curriculum generator with a different input modality. Where the standard flow takes a skill name and pulls reference footage from the web, Capture takes a 30-minute observation video and produces reference clips from inside it.

Capture is where the long-term moat sits.

Consider what Capture unlocks. A friend's grandmother making a family recipe. A barber giving a fade cut. A bartender making a signature cocktail. A mechanic diagnosing a specific engine. A professor walking a proof technique. Each capture is a curriculum nobody else has, because nobody else has that specific person. The library of captures becomes a public library of human craft, and the data trains the curriculum generator to be smarter at every skill on earth.

This is the same data flywheel logic that made Reddit, Stack Overflow, and YouTube into permanent properties. The user-generated knowledge accrues, the platform grows by Metcalfe's law, and competitors who arrive late cannot catch up by spending money. The bet on Capture is the bet on the flywheel.

When wedge logic still applies

Honest caveats. Wedge-free strategy is not always right.

When the content generation is still expensive in a domain, wedge logic returns. If you're building a coach for a skill where reference footage is rare or restricted (think: surgical procedures, niche industrial machinery, legally regulated practices), the agent can't generate the curriculum cheaply, and you're back to producing content per category. In those domains, pick a wedge.

When the user trust requirement is asymmetric, wedge logic also returns. A user trying to learn juggling is forgiving if Maestro's curriculum is suboptimal; the cost of a bad lesson is "I didn't quite get it today." A user trying to learn a regulated skill (CPR, firearm handling, prescription dosing) is not forgiving. In those domains, you need a wedge so you can verify the content with a domain expert before launching.

When the platform has to defend specific reference content (think: licensed sports moves, copyrighted dance routines), wedge logic returns. The content cost is high because the rights are expensive, not because the production is.

Maestro avoids all three by staying in the long tail of unregulated, observable, openly-referenced physical skills. Card magic, juggling, harmonica, sourdough, knot tying, latte art, calligraphy, pen spinning. The wedge-free strategy works here, and the wedge strategy is what dies.

The risks that could kill this

Form-grading quality varies by skill. Hand-close-to-camera skills (card magic, harmonica fingering) grade well with MediaPipe and Sonnet reasoning over landmarks. Full-body skills (gymnastics, surfing) grade poorly because the phone camera can't see enough of the body well. The mitigation is to lead with the skills the camera handles, and let AR glasses fix the rest later.

The first generated curriculum will sometimes be mediocre. If a user's first lesson on a new skill is wrong or boring, they won't come back. Mitigation: bias the first lesson toward extremely conservative, slow, "you cannot fail this" drills. Win the first session, earn the right to be ambitious in the next.

Capture-mode curricula could be bad enough to break trust. Generating a 14-lesson curriculum from 30 minutes of footage is hard, and the failure mode is "the agent missed the actual technique and produced a curriculum that looks plausible but doesn't teach." The mitigation is human review on the first N captures (you, watching the generated curriculum and editing the parts the agent got wrong) until the rate of "agent missed the actual technique" drops below about 10%.

Generalization beyond Maestro

The pattern generalizes anywhere the cost of new content categories collapses to near zero. Old playbook: pick one vertical, win it, expand. New playbook: build the engine, demonstrate breadth, let the dynamism be the demo.

This applies to learning, but also to:

• Coaching products of any kind. Same logic as Maestro for any skill where the lessons can be generated.
• Planning products. Any vertical where the planning logic can be expressed once and applied across categories. Leggo, the friend-group hangout agent, is a version of this.
• Content discovery. Per-user personalized exploration where the categories no longer need to be hand-curated.
• Tool generation. An agent that builds a tool on the fly per user request, instead of you building ten pre-built tools.

If you are starting a consumer product in 2026 and your first move is to pick a vertical to nail, the question to ask is whether the vertical is still the constraint or whether you are solving an outdated problem.