Improving the UX of (command-line) tools | Hacker News Reader

Improving the UX of (command-line) tools

blog.opalang.org

55 pointshbbio13 comments

Opa has some brilliant design decisions. Also, with the advent of Source Maps in Chrome and Firefox, the time is right for an end-to-end integrated web development language. Though I do find it strange as a long time OO person that "tightly coupled" is used as a positive trait in the "Hello Opa" post.

As I predicted years ago, Static and Dynamic typing are starting to lose their distinctions. I wonder if someone has started producing an as you type type inferencer? Given the right language design, I don't see why the editor can't keep track of all of these details, revealing types on mouseover and greying out vars that are ambiguously typed, much as syntax errors are hi-lighted in programmer's editors.

Already exists, e.g. F# in Visual Studio. Note however that type inference and dynamic typing, although similar on the surface in that you don't have to write type annotations, are completely different things.

The OO tenet says "Don't couple unrelated things". The converse of this is "Don't decouple related things", which is equally valid yet often ignored. Decoupling X and Y means that in the cases where you want to use X and Y together you need to connect them together again. This is a bad thing if in 99% of the cases you want to use X and Y together.

An example of this is the Java file API. In an effort to decouple a reference to a file, a reference to a stream of the file's contents, a reference to an arbitrary stream, reading from a stream, and reading from a stream in a buffered way, we end up with FIVE classes (File, FileInputStream, DataInputStream, InputStreamReader, BufferedReader) that will almost always be used together simply to read the contents of a file. This is of course an extreme example. A better design is to "couple" these things together in a gimmeTheContentsAlready(filename).

Likewise, the Opa guys are expecting that most people who use their stuff will use their database interface, and their language on both the server and the client, their easy client/server networking, and their http server. Shipping these as separate components would induce a lot of work to integrate them together again.

The OO tenet says "Don't couple unrelated things". The converse of this is "Don't decouple related things", which is equally valid yet often ignored.

Orthogonal areas of concern that are used together are just that: used together. They can be used together and packaged in a way that's super convenient. That still doesn't mean I want them tightly coupled.

NB: Works well together != tightly coupled.

An example of this is the Java file API. In an effort to decouple a reference to a file...we end up with FIVE classes

Yeah, I freakin hate that too. This Java JMI Api I used once, I had to instantiate seven entities from as many different classes just to get to send a message. But not all Java libraries are paragons of good object design. (You can leave off the Java, actually)

A better design is to "couple" these things together in a gimmeTheContentsAlready(filename).

Sorry, but you're mistaken here. Things are just as decoupled in the VisualWorks Smalltalk class library. There are equivalents to: (File, FileInputStream, DataInputStream, InputStreamReader, BufferedReader) and much more, but you can also still do:

    'file.dat' asFilename readStream

If you wanted to, you could just add a method to String named gimmeTheContentsAlready that would just return you the whole contents of the file with that pathname as a String. (I'd name it fileContents, though.) Either would take 3 short, easy to read lines of code, or one messy one, and about 2 minutes to write and save in your image.

Convenient methods and well designed APIs aren't the same as "tight coupling." If you think they are, then you need to read some better OO code.

EDIT: Another example -- at the same place there was a Sun message bus we used in the same product line, same functionality as JMI, just with better performance. Needed to instantiate 2 object to send a message with that.

It just depends on your definition of tightly coupled. By your definition Opa's implementation isn't tightly coupled either (the http server doesn't depend on the database, etc.). A better term is probably tightly integrated, and as far as I can see they also use this term on the rest of the website.

It just depends on your definition of tightly coupled.

Back in my day, "tightly coupled" was somewhat like the opposite of "modular."

I believe f# type providers are an implementation of what you're describing: http://research.microsoft.com/apps/mobile/video.aspx?id=1500...

At the risk of sounding like a broken-record aging neckbeard (reality: I'm a college student with light stubble), I think the title of this thread is very misleading, even if the blog post itself contains valuable material.

First, I'm just barely old enough to remember using DOS on a machine without Windows -- which I believe is the last time that a GUI/DE wasn't viewed as 'essentially essential' for an end-user computing environment. However, I'm just young enough to remember growing up with Windows for most of my computing life, so it's not like I was indoctrinated into the ways of the command line from an early age. Far from it - I'd forgotten all the DOS commands I knew by the time I turned 10, and I actually only began learning Unix very recently (while in college).

I don't think command-line tools have a UX problem. If anything, we have a teaching problem. Command-line tools have a slight learning curve, but really nowhere near as much as many people think. The story of how I got into using the command-line for all (well, almost all) my daily tasks is a bit too long to post here (unless people are interested), but in short: I went from not knowing the difference between cd and dd all in the span of a month. (Thankfully, I did not have to learn those two commands the hard way!)

Why was I able to do it? Because Unix has a ridiculously uniform design, and is unbelievably modular. UI designers today should strive for the uniformity that my GNU system has. (I use GNU, but this comment applies for the most part to other systems, like the BSD-based varieties, and is largely, though not completely, true even when comparing two different systems!). There are a few wrinkles around the edges, but when you consider the very distributed/patchwork way that modern POSIX systems came to be, it's almost miraculous that the single-dash switches vs. double-dash options distinction works in most cases the way you'd expect. The difference between required arguments and optional parameters is implemented very well in most utilities (it's rare that I have to specify unnecessary flags for required parameters, for example). Combining switches also works the way you'd expect on most utilities, and even the naming is consistent (-h, -v, -l, -a, --help, --verbose, etc. follow the Principle of Least Surprise (POLS) in all but very few cases). Even if there are differences between, say, ls -G on BSD and ls --color on GNU, that's a matter of the way those basic utilities were implemented, and the interface itself is still fairly uniform. (And I'm willing to grant some leeway for such fundamental utilities as 'ls', since these tools develop over a period of time).

That's not to say that everything has to follow the Unix-like model. I'm not saying that Opa's efforts aren't valuable simply because they're not standard POSIX. On the contrary, I think that tailoring a command-line tool to the task at hand can be very valuable - SQL queries fall into this category as well.

But they're valuable because they provide a modular (scriptable) tool that modify a familiar interface to make it easier to use in a domain-specific context with minimal violations of the POLS. That's fundamentally different from improving the UX of the command line; instead, it's applying the general UI principles of command line tools to a specific context. All instances of a general UI violate the POLS in some way by definition; the goal is to minimize that.

In the end, I can't think of any other UI that follows the Principle of Least Surprise as well as Unix does, and it's the most fundamental command-line tool in my life. We may have a problem with the command line, but the command line does not have a problem with UI.

(Edit: Does someone know how to write an asterisk without causing it to be parsed as the opening of an italicized block, as in <asterisk>-nix to refer to the wildcard 'Unix-like'? I'm on choppy in-flight internet and can't figure this out reliably with trial-and-error).

I think there is much more that can be done to make tools more user-friendly.

An example that pops to my mind is when you misspell a git command, and git suggest what you could've intended:

    $ git pill
    git: 'pill' is not a git command. See 'git --help'.

    Did you mean this?
        pull

That's already one step towards a better UI, but it could go even one step further if I could answer the question, ie: I press 'y' and git will pull.

Of course I'm not criticizing GIT (which I think as a better UI than many tools already), it's just an example.

zsh will actually do this for commands if you make a typo when the correct option is turned on. I don't think it will correct options though (but can autocomplete).

I'm not saying that the UI couldn't be improved - I'm just saying that what they're doing isn't really an improvement on the the command-line as a UI, as it's just a domain-specific application. (And, I implied this but didn't state it explicitly: domain-specific applications should aim to be as 'compatible' as possible from a UI standpoint with the generic tool - ie, Unix - deviating only when there's a specific need to, and defaulting to the generic toolset. I'd need to see more to judge if Opa fits this; it's just a general principle).

More importantly, though: what you're looking for can already be done! Just use zsh (or, even simpler: a bash correction tool).

Because Unix is so modular, you can make these kinds of UI changes without breaking or even changing the underlying toolset. That means that you and I can tailor the UI to our specific needs, and we don't even really have to touch the general purpose toolset. Other examples include using Vim keybindings for the shell instead of Emacs (try modifying keyboard shortcuts with most pre-compiled GUI applications!) or scripting and aliasing.

Thanks for mentioning that, actually, because it illustrates my point very well! The command-line UI's strength is that it's literally as modular and as minimal as possible (at least, as far as I can conceive of). That's what I personally love the most about it. It defines only the bare minimum of tools necessary to satisfy two conditions:

1. By default, [almost] everything works as expected, with no customization (but assuming knowledge of the system). 2. Those who want to customize will find that it provides all of the tools necessary to create the exact interface they want, without modifying the underlying toolset as well. (This customization can be done to an essentially arbitrary level -- the technical limitations are well beyond what a reasonable person would consider 'sane').

You may make the argument, then, that the command-line isn't even an interface at all! It's more like an abstract interface that lets you stitch the tools together to define the interface you want, all the while providing a 'default interface' - a sensible set of defaults for those who can't be bothered to customize.

(Edit: I understand that tools aren't perfect - the git-pull/git-push asymmetry is an example - but the inconsistencies are vastly outnumbered by the overarching uniformity. And again, with such a modular model, you can create the UI you want without waiting for Linus to change the behavior of git).

Another example of catastrophic UX is certainly the unix command 'find', as in: find . -name "*.opa" -exec cat \{\} \; | wc -l

I wonder if any Unix beginner has ever managed to find the syntax for -exec without copy-pasting the examples :)

One aspect of having composable components is knowing how they get composed. The fact that {} and ; may need to be escaped is something having to do with the shell, not with find and it's -exec syntax.

In fact, this is one of those things that is obvious once you learn about shell and how commands lines get expanded and executed. The problem there is that most people don't learn about the shell until forced to by being bitten by something goofy and unintuitive like -exec arguments. The order in which people learn about these things, and it being difficult to discover (in that I can suggest a way to "verify" the command line before running it, but that doesn't help if you don't know how to interpret the results, which also requires experience dealing with the system as a whole) is what causes the catastrophic failure.

That these may need to be escaped is even documented in the find(1) documentation.

    -exec command ;
    Execute command; true if 0 status is returned.
    All following arguments to find are taken to be arguments to
    the command until an argument consisting of `;' is encountered.
    The string `{}' is replaced by the current file name being
    processed everywhere it occurs in the  arguments to the command,
    not just in arguments where it is alone, as in some versions of
    find.  Both of these constructions might need to be escaped
    (with a `\') or quoted to protect them from expansion by the
    shell.

Let's also not forget that the whole flags/arguments debate is often not really that useful in todays *nix command line, either. I'd love to see more command-line utilities work like the Heroku gem and less like sed. (Yes, I know it breaks the whole Unix philosophy of one tool == one action, but from my experience it is far easier to get new people to learn.)

Reply on news.ycombinator.com