|Perl: the Markov chain saw|
GUIs Considered Harmful by Tom Christiansenby princepawn (Parson)
|on Sep 12, 2001 at 09:19 UTC||Need Help??|
This short article by Tom Christiansen is nothing short of a classic. I italicized the part that is most important to me. This was written about 10 years ago (1991 or so) and posted to usenet.
GUIs Considered HarmfulI am increasingly troubled by how many new applications are designed to work solely under a GUI. While this may make some amount of sense for people coming from the PC or Mac worlds, one of the strengths of Unix has always been the ability to use it from anywhere. These people don't seem to understand this.
Of how much ultimate utility is that nifty new spreadsheet, editor, or debugger if I can't dialup from the field and run it on my vt100? Too often a tool that "does windows" is little more than a marketing gimmick to dazzle impressionable users into not noticing that they don't have the real functionality they need.
GUI-minded programs seldom lend themselves to being used as components in larger tools. As such, they do not fit well into the UNIX tool-and-filter philosophy. Instead of each being a single program that modestly attempts to do one thing well, they are a throwback to the Bad Old Days when each program was a standalone, monolithic monster that didn't interface with anything else.
It's all well and good to place a GUI wrapper around an existing tool, but to design a new application with only a GUI interface in mind is to forever limit that tool's flexibility. After all, how to you write a shell script that drives an automated xrn session?
Providing programmability for the fancy graphics software remains an open problem. The most effective use of GUIs in UNIX environments is to design the nitty-gritty computational function as a "back end" that can be driven either manually or automatically.
The GUI wrapper should be a separable module. If they're plug-replaceable, the application isn't irretrievably wedded to any specific GUI technology, such as SunView, NeWS, or even X11 or its children, like Open Look or Motif. Sending standard commands down a pipe the way the STDWIN or wafe packages behave is also a reasonable approach.
This means your program should be runnable both with and without the GUI present, and accept input from a mouse or under programmed control. Preferably that means both a shell-level interface for convenience and a C-level interface for efficiency; Perl programmers could elect either method. That way, naive users can use push-button GUIs, blind users can use Braille terminals, and sophisticated users can program solutions to intricate problems.
It has been noted that GUIs make simple things simple, and complex ones impossible. Certainly it is worthwhile to make simple things simple. But too often software is geared to only one level of expertise. That which is novice-friendly is too frequently expert-hostile, and vice versa. Being needlessly forced to click the mouse over a menu will slow down the expert user who is more comfortable with a keyboard interface.
Gratuitous distractions from the keyboard only slow down the experienced user. A precision pointing device that didn't require taking your hands off the keyboard would help. There are cases where only a GUI makes sense, like a CAD system. Being able to delineate a region or draw a figure with a mouse is probably a reasonable use for it, but selection of a range of possibilities isn't, at least not after you've become familiar with the tool.