Predictive text FTW

I am just starting to use Gnome Do, and it’s amazing. You hit (windows key) + space, and it pops up a little window where you give it commands. It tries to figure out what you mean when you type a certain combination of words, and remembers what you usually do on those combos, and does that in the future.

Eventually this could replace my entire need for a command line, in a much more intuitive and user friendly sense. Getting instant feedback, and seeing where you can stop typing into the command gives you a “free hint” as to the minimum amount of typing required to perform an action.

Presumably all actions can’t be condensed down into a simple combination of 3-4 keystrokes (the level where you see a noticable time savings), so only the most commonly used actions will be useful for this (the whole learning from you thing). I believe that if somehow you incorporated context into the choice when you are figuring out what to call from what combination of letters, you could solve that problem….The hard part, defining the context.

I think that for example, since I’m writing a blog post, and I type (after calling Gnome do) ‘sp’, then I want to check the SPelling of whatever I’m typing. However, if I’m inside my mail application (or G-mail if you want to get fancy), ‘sp’ means SPam, either that this is spam, or call the spam dialog (or whatever I usually do when I type spam(while viewing a message, or from the Start screen)). This was a trivial example, but it lead to a very easy analogy. I feel that this concept could be extended almost indefinitely, with the only limitation being on how we define context, and how smart the algorithm is.

If the algorithm gets good enough, you get to stop typing the shortcut.

Edit: I feel this is the same way that spam filtering currently works. Look at all of the incoming mail, and look for patterns in it. If you dont’ define an “action” and a “recipient”, then the input becomes one item, dependent on the context. I think you could adapt a Bayesian algorithm to do this very thing. (Completely ignorant on how Bayesian filters actually work, could be way off point).

I also believe the major downfall of this idea is the concept of “modes”. Modes are generally a bad thing, and having all of this context will confuse some people. If they don’t understand the different contexts, and doing the same thing has different results sometimes, they will consider it “broken”. I think the visual feedback provided by the program will be crucial in how this is interpreted. Having the feedback will allow the person to know what the program is going to do. Also, some visual representation of the current context (like a favicon in the corner) will allow them to understand what context their in, and why the program is doing what it is doing.



Hey there! I'm Eric and I work on communities in the world of software documentation. Feel free to email me if you have comments on this post!