They are live now on the Amazon Alexa "skill store". Go enable them (direct links below), try them out, and let us know what you think. We'd love to hear any and all feedback.
CompliBot: http://alexa.amazon.com/#skills/amzn1.echo-sdk-ams.app.cb3e2030-f6a7-43f3-83da-164cccd07896/activate
InsultiBot: http://alexa.amazon.com/#skills/amzn1.echo-sdk-ams.app.2fb024ea-b485-4966-be4e-1a42f26ca9ef/activate
(Note: You have to have an Alexa capable device tied to your Amazon account to see the skill store)
The Story of Their Creation
That said, the whole point of these was to get our foot in the door and learn the system, and we certainly did that. As an added bonus, we got a full-fledged framework out of the process that should help accelerate some of the more advanced projects we have in the pipeline. And we did end up with a few things that we think make our skills unique relative to what else is out there:
- Targetable quips: In testing, we quickly discovered that users often wanted to do things like ask insultibot to "insult David". To accommodate this, we have a subset of our quips that can be targeted at individual users, and the sentence structure will be modified on the fly with the name given.
- Bots with personality: We wanted our bots to have character beyond just the simple insults and compliments, so we made sure all of our conversation flows - "help", "who built you?", etc - were spoken the way someone super friendly or super disdainful would speak.
- Quip tracking: In order to make sure you weren't hearing repeated quips, or even repeats of certain patterns with different punchlines, we now keep track of which quips you've heard during an ongoing session with one of the bots. In most cases, we won't repeat anything you've heard in your last 10 requests!
- Gradual prompt backoff: As part of our certification process, we were asked to add the questions you hear at the end of each statement ("Would you like to do more?"). We weren't fans of this approach, but the certification team was adamant, so we ended up compromising with what is essentially a declining prompt the longer you maintain your session, until eventually it's not actually asking you to take actions, since presumably you already know how. As far as we know, we were the first skill to be certified with this compromise solution (in fairness, the idea wasn't ours - props to jjaquinta for that).
- Unique content: All of the quips we have published came directly from the minds of DERP or our friends (with the exception of a single nod to "The Simpsons" in CompliBot).
- Easter eggs: There are a couple neat things tucked away - we'll see how long it takes people to discover them. =)
We also had to take out or pause a ton of cool ideas. We don't want to spoil everything, since we may sneak some of these ideas back in later, but a few of the things that got pulled or scrapped for certification or voice recognition reasons are:
- Random conversation: This goes hand-in-hand with the first bullet above - wanting our bots to have real personality - but at one point we had implemented some random smalltalk.
- Topical quips: Think "give me a Star Wars compliment" or "insult me about sports".
- User specific quips: As the bot gets to know you, it's better able to target you.
In the end, we decided to just focus on the core competency of the bots to get things out the door, and given the certification problems we ran into (despite the simplified approach) that was probably a good call. We're not gonna harp on that in this post (but you can read the articles here and here) , but suffice it say the last month has had its share of rage. All's well that ends well, though, and now we're live and serving up insults and compliments.
Go check us out, and leave us a comment if you have any feedback about pretty much anything. We'd love to hear your suggestions for features, or even for new quips.
-The DERP Group team