Dustin Coates in Reinvent

AWS re:Invent 2018 Days Four and Five Recap

Here are notes from days four and five of AWS re:Invent 2018. If you missed day one, you can read those notes here, or read day two or day three. Want to continue the conversation? Reach out at dustin@dcoates.com or on Twitter.

Deep Dive into Three Critical Aspects of Alexa’s In-Skill Purchase API

Thursday started with another session on ISPs (in-skill purchases), hosted by Mark Kockerbeck and Muhammad Yahia, both on the engineering management side for Alexa. Kockerbeck and Yahia took a Halloween-themed skill (to ride the seasonal traffic patterns) and experimented with different monetization schemes.

They said early on in the session that ISPs are “the gate around the content, not the content itself.” The customer then pays to remove the gate. Honestly, I don’t know how the difference between the gate and the content matters. My closest interpretation is that the gate-content distinction is important because Amazon won’t allow a “premium-only” skill, or a skill that users must pay to interact with at all.

There are three types of digital in-skill purchases:

  • One-time
  • Consumable
  • Subscription

The two added one-time and consumable purchases to the Halloween skill. What they discovered was that people should know exactly what they’re going to receive, people need to be in a purchase mindset before the skill tries to upsell them, and the skill needs to back off on the upsell if the user expresses disinterest.

This might seem intuitive, but what was counterintuitive is the number of steps in the buying process. Users need to express interest within the skill, and then affirmatively purchase from Alexa. The team began with a single step, but found that it threw too much information at the user at once. The two steps added more friction, but each step was small enough to be digestable. It ultimately improved conversion rates.

While skill builders do not find out if there was a refund other than a sudden change in the user’s entitlement, users need to know what they bought and what they can buy. The help intent is a useful place for this. This was the only place where users could purchase a “werewolf” upsell and it was quite successful.

A technical note that doesn’t fit anywhere else, but I still found interesting: the skill session ends when the buying process begins and session attributes are lost.

Three Lessons from “Escape the Room” That Apply to Making Money with Your Alexa Skills

Paul Cutsinger of the Alexa team, joined Gal Shenar, skill builder and found of Stoked Skills. Shenar has built multiple skills, include a series of escape room games. His experience is useful for game builders, but also for more builders of more “serious” skills.

Because sessions were repeated, the presentations from the embedded videos might not match exactly what I present here. This is the summary of the session that I attended.

Experimentation is “huge” when building skills. Start small and then expand. Shenar was interested in experiences that could take hours, and came to the escape room idea. He didn’t feel there were many immersive games at the time that people could play together for an extended period of time, and he wondered if he could shrink an escape room to voice-only. Still, that idea and all ideas must evolve. If you start by designing the story and creating the personae, the skill will evolve during those phases, just like it will during testing and after launch.

If you take that knowledge, you’ll then move to modeling the intents. Escape the Room is both open-ended, and surprisingly low on the number of intents. Shenar says this is on purpose, because he wanted users to know what they could do next. Therefore, they can only look around, look at objects, use items with objects, or solve puzzles. During the development process, he actually shedded a fifth intent, picking up items, because it wasn’t useful and prevented focus.

Shenar developed the skill by separating the content and the code (or, as Cutsinger put it: a game engine and a level editor). All of the responses are gathered in a single place, and the logic is in another.

The response creation process, and the analytics of usage, lead to insights itself:

  • Shenar started by adding a longer intro, but people dropped
    • Cutsinger provided the “one breath” guideline, where most responses should be sayable by Alexa in a single breath
  • Didn’t provide guidance as much, because users were leaving negative reviews that the corrections seemed “naggy”
  • Added variety of responses, plus sound effects, because they “convey a lot of [imagery] in a little time”

To conclude, there were three main takeaways:

  1. Easy to play, hard to master
  2. Help should appear when needed, and disappear when done
  3. You will not get it right the first time; observe and adapt the difficulty level