Hello Alexa

Posts in This Series

  1. you are here  


Here's a post I've been working on for longer than I care to admit to. For a few months, I've been wanting to kick off a series on serverless development practices, theory behind why serverless is such a game changer, as well as a couple of demos. The fact of the matter is, each time I've started, I've gotten so bogged down in writing a post about what serverless is and the backing theory, I've managed to loose blogging momentum and it has since idled in my drafts folder for a while. So, what I've settled on for now, is a brief introductory to my first Alexa Skill, and the basics of how it works.


Have you gotten past your fears of an always listening device in your home? Has conversational computing started to change how you think about software development? Have you wanted to build your first Alexa Skill, with more than a simple "hello world"?

Amazon's Alexa (Echo) devices and services, along with Google's Home line of voice/audio products, are changing the way everyday computing exists and is used in our lives. They're not the only ones, along with Siri, Cortana, and more, these voice and audio services have helped to begin a bit of a minor revolution. This one doesn't require singing and dancing, unless that's your thing, but rather that we re-examine the nature of how we interact with computing devices. The very nature of being able to speak aloud, give instruction in a natural langauge format, and receive information is probably one of the more underrated paradigm shifts of this decade. With any luck, it will only go onwards and upwards from here... unless Skynet takes over.

A while back I tweeted out a short poll on what sort of Alexa Skill I should build. I included a couple options and got a couple votes back. Here are the results.

This will feed the next post, which covers the Alexa Skill I built. For now, let's have a look at how a basic Alexa Skill is structured.

Constructing an Alexa Skill

At the most basic level, an Alexa Skill is comprised of three main parts:

  • a backing service, generally a serverless implementation, hosted somewhere that the Alexa skill can invoke

    • this can be a publicly accessible HTTP(S) endpoint
    • for Amazon's product usage, the most common is AWS Lambda
    • AWS Lambdas use their ARN (Amazon Resource Names), which is an Amazon internal name space; quite easy to talk from AWS Lambda to an Alexa Skill (point at the right ARN)
    • for Alexa skills, most often make use of the Alexa SDK or related packages (to assist in handling the hooks in a consistent fashion)
  • a configuration of the Alexa Skill

    • matches vocal commands such as name and actions, responses
    • it matches registered phrases to "intents"
    • these intents can be either standard ones from Amazon, such as "cancel", or custom ones such as "do this thing I defined"
  • it is published to the Alexa Skill "store" on Amazon

    • it can be available to test pre-release / while in development, for the developer's associated account (and beta testers)
    • the publishing process involves review
    • my experience has been around 24 hours turn around
    • some things are annoying in that they can be quite arbitrary

      • like one of my skills which fit their described naming conventions as it is listed, but their review said it didn't; they were empirically incorrect, but there is no appeals process

How About an Example Skill?

Before we go further, I'd like to ask you to think about how we register human speech and match up the intentions behind them to any conveyed ideas or actions to take. It's no small task, but thankfully we don't have to reinvent the wheel. For Alexa Skills, the Alexa Skills Kit brings a fair amount of tooling easily installed via npm (and respective equivalents for other runtimes and languages). For the most part, we'll have our functions that define what happens when an action is taken, and the "Intent" definitions, that call corresponding actions. There's lots of room to talk about what these mean and I am taking a somewhat simplistic approach to the theory, but enough theory for now! I know I'm bound to come back to the theory, I seem to do so more than I might like, but for now, let's assume we know what we're doing and why. On to the examples!

Example 1: Hello World

Stop me if you've heard this before, but we'll begin with a good ol' "hello world", you know this one, right? So, what do we need to start with?

Scaffold The Project
  1. this is a Node.js example, so we'll be creating a new project directory and initializing with npm

    • mkdir hello-world
    • cd hello-world
    • npm init -y
  2. we'll use the official Alexa Skills Kit SDK for Node.js

    • npm install --save alexa-sdk
  3. we'll have a short and sweet Node.js module, so create the source file

    • touch index.js
  4. massage your package.json, update your:

    • name
    • description
    • ensure the "main" points to index.js (that should be the default)
Build Out index.js

To correctly hook into the epxected events, we'll export a handler function, containing hooks to our various "intentions". This is what is expected by the Alexa SDK; for more on registering handlers, check out their wiki.

Here's a start to this file:

'use strict';
const Alexa = require('alexa-sdk');
// this exported handler is what defines how the Alexa runtime can interact with our module
exports.handler = function(event, context, callback) {
const alexa = Alexa.handler(event, context); // a new instance
alexa.registerHandlers(handlers); // registers handlers (below)
alexa.execute(); // now it's ready
// we'll populate these shortly
const handlers = {}
view raw begin.js hosted with ❤ by GitHub
Configure Your Handlers

Since we are now bringing in our handlers, here's a quick look at what is entailed. For starters, it's an object, as you can see from the declaration above. We will match our intent names to a function, which will perform actions and emit an action based on some pre-defined actions. These can also be a different registered intention, such as chaining an "I didn't understand that" prior to calling the help intent response.

Here's a basic configuration for "Hello World".

const handlers = {
'LaunchRequest': function () { // just opening the skill
'HelloWorldIntent': function () { // the main intent
this.response.speak('Hello World!');
}, // now begin wiring of canned intents
'AMAZON.HelpIntent': function () {
const speechOutput = 'This is the Hello World Sample Skill.';
const reprompt = 'Say hello, to hear me speak.';
'AMAZON.CancelIntent': function () {
'AMAZON.StopIntent': function () {
this.response.speak('See you later!');
'Unhandled': function () {
view raw handlers.js hosted with ❤ by GitHub

To break things down, you'll notice there are two forms of intents, custom and standard Amazon intents. The Amazon standard intents are common across skills, and cover cancellations, stopping, and asking for help.

The first custom intent is the LaunchIntent. That registers an action, "speaking" a phrase and emitting an event.

Next Time: A Better Example

I have a better example for the next post, the skill I wrote following the above twitter poll. Next time I will cover an Alexa Skill I have published, along with some show-and-tell about configuring the skill with intentions in the AWS Lambda and Alexa Skill Consoles.