Because painters leave a trail of work behind them, you can watch them learn by doing. If you look at the work of a painter in chronological order, you'll find that each painting builds on things learned in previous ones. When there's something in a painting that works especially well, you can usually find version 1 of it in a smaller form in some earlier painting.

I think most makers work this way. Writers and architects seem to as well. Maybe it would be good for hackers to act more like painters, and regularly start over from scratch, instead of continuing to work for years on one project, and trying to incorporate all their later ideas as revisions.

The fact that hackers learn to hack by doing it is another sign of how different hacking is from the sciences. Scientists don't learn science by doing it, but by doing labs and problem sets. Scientists start out doing work that's perfect, in the sense that they're just trying to reproduce work someone else has already done for them. Eventually, they get to the point where they can do original work. Whereas hackers, from the start, are doing original work; it's just very bad. So hackers start original, and get good, and scientists start good, and get original.

The other way makers learn is from examples. To a painter, a museum is a reference library of techniques. For hundreds of years it has been part of the traditional education of painters to copy the works of the great masters, because copying forces you to look closely at the way a painting is made.

Writers do this too. Benjamin Franklin learned to write by summarizing the points in the essays of Addison and Steele and then trying to reproduce them. Raymond Chandler did the same thing with detective stories.

Hackers, likewise, can learn to program by looking at good programs—not just at what they do, but at the source code. One of the less publicized benefits of the open source movement is that it has made it easier to learn to program. When I learned to program, we had to rely mostly on examples in books. The one big chunk of code available then was Unix, but even this was not open source. Most of the people who read the source read it in illicit photocopies of John Lions' book, which though written in 1977 was not allowed to be published until 1996.

Another example we can take from painting is the way that paintings are created by gradual refinement. Paintings usually begin with a sketch. Gradually the details get filled in. But it is not merely a process of filling in. Sometimes the original plans turn out to be mistaken. Countless paintings, when you look at them in x-rays, turn out to have limbs that have been moved or facial features that have been readjusted.

Here's a case where we can learn from painting. I think hacking should work this way too. It's unrealistic to expect that the specifications for a program will be perfect. You're better off if you admit this up front, and write programs in a way that allows specifications to change on the fly.

(The structure of large companies makes this hard for them to do, so here is another place where startups have an advantage.)

Everyone by now presumably knows about the danger of premature optimization. I think we should be just as worried about premature design – deciding too early what a program should do.

The right tools can help us avoid this danger. A good programming language should, like oil paint, make it easy to change your mind. Dynamic typing is a win here because you don't have to commit to specific data representations up front. But the key to flexibility, I think, is to make the language very abstract. The easiest program to change is one that's short.

Figure 2-1. Leonardo's Ginevra de' Benci, 1474.

This sounds like a paradox, but a great painting has to be better than it has to be. For example, when Leonardo painted the portrait of Ginevra de' Benci in the National Gallery, he put a juniper bush behind her head. In it he carefully painted each individual leaf. Many painters might have thought, this is just something to put in the background to frame her head. No one will look that closely at it.

Not Leonardo. How hard he worked on part of a painting didn't depend at all on how closely he expected anyone to look at it. He was like Michael Jordan. Relentless.

Relentlessness wins because, in the aggregate, unseen details become visible. When people walk by the portrait of Ginevra de' Benci, their attention is often immediately arrested by it, even before they look at the label and notice that it says Leonardo da Vinci. All those unseen details combine to produce something that's just stunning, like a thousand barely audible voices all singing in tune.

Great software, likewise, requires a fanatical devotion to beauty. If you look inside good software, you find that parts no one is ever supposed to see are beautiful too. When it comes to code I behave in a way that would make me eligible for prescription drugs if I approached everyday life the same way. It drives me crazy to see code that's badly indented, or that uses ugly variable names.

If a hacker were a mere implementor, turning a spec into code, then he could just work his way through it from one end to the other like someone digging a ditch. But if the hacker is a creator, we have to take inspiration into account.

In hacking, like painting, work comes in cycles. Sometimes you get excited about a new project and you want to work sixteen hours a day on it. Other times nothing seems interesting.

To do good work you have to take these cycles into account, because they're affected by how you react to them. When you're driving a car with a manual transmission on a hill, you have to back off the clutch sometimes to avoid stalling. Backing off can likewise prevent ambition from stalling. In both painting and hacking there are some tasks that are terrifyingly ambitious, and others that are comfortingly routine. It's a good idea to save some easy tasks for moments when you would otherwise stall.

In hacking, this can literally mean saving up bugs. I like debugging: it's the one time that hacking is as straightforward as people think it is. You have a totally constrained problem, and all you have to do is solve it. Your program is supposed to do x. Instead it does y. Where does it go wrong? You know you're going to win in the end. It's as relaxing as painting a wall.

The example of painting can teach us not only how to manage our own work, but how to work together. A lot of the great art of the past is the work of multiple hands, though there may only be one name on the wall next to it in the museum. Leonardo was an apprentice in the workshop of Verrocchio and painted one of the angels in his Baptism of Christ . This sort of thing was the rule, not the exception. Michelangelo was considered especially dedicated for insisting on painting all the figures on the ceiling of the Sistine Chapel himself.

As far as I know, when painters worked together on a painting, they never worked on the same parts. It was common for the master to paint the principal figures and for assistants to paint the others and the background. But you never had one guy painting over the work of another.

I think this is the right model for collaboration in software too. Don't push it too far. When a piece of code is being hacked by three or four different people, no one of whom really owns it, it will end up being like a common-room. It will tend to feel bleak and abandoned, and accumulate cruft. The right way to collaborate, I think, is to divide projects into sharply defined modules, each with a definite owner, and with interfaces between them that are as carefully designed and, if possible, as articulated as programming languages.