Related
I was looking at some study i have to do in the future to do with procedural generation techniques and i was wondering what type of content you have:
Developed
Helped Develop
Seen implemented
Tried to develop
and what methods/techniques/procedures you used to develop it.
If you feel generous maybe you can even go into specifics of it such as data structures ad algorithms you have used to develop it.
If this needs to be put as community wiki because it is not me asking for a problem to be solved just let me know.
This is not a homework thread because it is a research unit that i'm not taking yet ;)
Introversion software, the makers of the games Defcon, Uplink and Darwinia (among others) have started working on a game about a year ago which extensively uses PCG for city generation, here is a video of their work, and you can read more about it on the development diary of the game (start from the first part at the bottom of the page!).
This immediately got me extremely interested, and seeing the potential for games I immediately started researching the technology. I have amassed a folder of 18 PDFs about the subject (research papers, SIGGRAPH presentations, etc). Here, I uploaded it for you.
The main approach is to use L-Systems, however, I never got around to understanding enough of that to make something out of this. I tried other, less successful approaches like using Voronois, recursively splitting a rectangular area into more smaller areas and shifting the boundaries a little to obtain a bit of randomness and polygon division.
The last method I had gotten from Mike's Code Blog's posts (here and here). The screenshots shown on his blog make me drool, it is my biggest programmer's dream to ever get something that looks like that. I emailed him to ask how he did it, and here is the relevant part of his reply, I'm sure he wouldn't mind me posting this here:
L-Systems is definitely one way to go, but that isn't what I'm doing. The basis of my method is polygon subdivision. I start with a simple polygon that represents the entire area of the city. Then, I split it (roughly) in half, and then split those two polygons, etc. until I get down to city-block size. At that point, the edges of all my polygons represent roads. I then use the same subdivision method to break the blocks down into building-size lots.
The devil is in the details, of course, but that is the basic method.
I for one still haven't managed to fully implement a solution of which I'm satisfied of, but it remains one of, if not my single biggest programmer's dream to ever achieve something like this.
Here are a few of the leaders in procedurally generated terrain (and to a lesser extent foliage). If you don't get a detailed answer here regarding methods and techniques, you might want to look in / ask in their forums. I have seen some discussions of techniques there.
TerraGen 2
World Builder
World Machine
Natural Graphics
Noone mentioned the demoscene that ONLY use procedural stuff?
So, go search for Werkkzeug, Kkrieger, MilkyTracker to start. Also you can visit the site pouet and see the wonder of well done procedural videos (yes, procedural videoclips! With music and graphics, all procedural!)
Allegorithmic's products are used in actual shipping titles. These guys focus on texture generation (both offline and at runtime).
They have some very pretty screenshots and demos.
When i start writing something complex, I find that restart the writing like 10 times before I end up with what I want, often discarding hundreds of lines of code.
Am I doing something wrong, or do others have workflows like this?
EDIT: Right now, I'm working on a modular compiler. The last project I was working on was a server in java. Before that it was some concurrency stuff.
I do a fair bit of planning, and I never start coding before I've got interfaces for everything.
Given that, is it normal to just wipe the slate clean repeatedly?
Discarding many lines of code is usually a positive aspect of refactoring. That's great. But starting over ten times means that you probably haven't analyzed your problem and solution. It's fine to backtrack and sometimes to start over but not that often. You should lay out your code in such a way that when you backtrack and refactor, you keep most of what you created because it will exist in nicely isolated and logical chunks. (Using vague language since language of choice wasn't specified.)
From author's comment:
Usually I restart because I get
confused by all the stuff going on in
my code.
Study your craft and make good use of design patterns and other best programming philosophies to lend your code a well-defined structure... something you'll recognize months and even days down the road.
If you're starting something complex, a little planning before you start writing would seem to be a good idea.
Design first.
Perfectly normal. No matter how much I plan ahead, I very often have an "Aha!" moment once the hands hit the keyboard.
Just as often it's a "What the heck was I thinking?" moment.
It's all good. You're making better code.
All the suggestions here are valid, however, remember that there's a moment in a programs lifetime that is "good enough". It's easy to fall into a trap of never ending refactoring, just because you see that "yes, this could be done better!". Well, face the truth -- unless your program has just a few lines, there's ALWAYS a way to do it better.
I believe there are happy programmers out there that don't suffer from that, but at least I need to keep reminding myself that there's a line that's called "good enough".
And it's especially true if you're writing code for someone else -- nobody will note that you did something "nicer", all that counts is "does it work well?".
Also, a VERY GOOD practice is at least to get it to WORK before rewriting. Then you can always fall back to a working previous solution.
(since 12 years I'm constantly rewriting a game I'm writing, and I'm nowhere near the end...)
On a complex problem, this is common. If you aren't totally stabbing in the dark, it really helps to sketch out our ideas first, but then again you're just moving the 'retries' from code to paper.
If it helps you get to a good solution, how can it be wrong?
It depends on how well I know the problem space. If it is familiar territory, then I'd be worried if it took 10 iterations. If it is unfamiliar territory, then it might take as many as 10 iterations, but at least some of them would be reduced to prototype - or an attempt at a prototype - before being discarded.
If you are learning something with each iteration then there probably isnt a problem. Deadlines and pesky things like that might make your life a little more difficult. :)
When I am working on a new problem I like to pseudocode it out in comments in the actual function handler, as part of generating the stub for my TDD. Then add the code in to each step I had in the comments of the function body.
It helps me to keep focused on what the problem that I am solving is, and not get lost in the details to early.
The single biggest change you can do to help yourself would be to plan your code first. On paper.
Your plan doesn't have to be super in-depth (although sometimes that's good too). Just sketch out a rough idea of what you want your program to do. Jot down the key functionality points. "I want it to do this, this and this".
Once you have that, sketch out a rough design. Classes, methods, functions, objects. Give it a little form. Do a rough allocation of functionality to various portions of your design.
Depending on the nature of the project, I might take a rough design such as that and turn it into a much more detailed design. If it's a small project, maybe not. But no matter the projected complexity, time spent designing will reward you with better code, and less time spent coding. If you have obvious mistakes that require you to refactor large portions of your program, they should be apparent in your initial design and you can adjust it. You won't have wasted hundreds of lines of code on an obvious mistake.
Compilers are very complex applications, and you can't write an optimizing compiler from start to finish in one pass - no matter how much thought you put into it at first. Usually you attempt to get something to work correctly from start to finish and then go back to modularizing it and adding new features like optimizations. This process means lots of refactoring and replacing whole sections outright. This is also part of the learning processes - as no one can know everything and remember it!
(I'm also working on a .NET compiler as part of the MOSA project - www.mosa-projet.org.)
solve your problem on paper . . . dont be in such a rush to type.
There are two situations. (1) I've been able to confidently plan ahead and isolated my abstractions. (2) I haven't.
For (1), an effective technique is to put in dummy versions of certain classes or functions just to drive the rest of the code. (or conversely, to write said classes and functions and drive them with a test script.) This allows you to tackle only part of the complexity in each pass.
As much as everyone says people should plan in advance, it often doesn't work that way, resulting in situation (2). Here, be careful to manage what you are trying to accomplish in one iteration of code. As soon as you find your brain unable to juggle all the things you are doing, scale back your ambition for what you want to achieve before the next compile-and-test. Allow your code to be flawed but easy-to-write on the first pass, and then develop it through refactoring. This improves efficiency over repeatedly wiping the slate clean.
For example, one way I used to get into messes was by sniffing out common code and refactoring into subroutines too early, before I really knew the shape of the code. I've since started allowing myself to duplicate code on the first pass, and then going back and factoring it into subroutines later. It has helped tremendously.
It's called refactoring buddy and it's good, you just need to limit it so you won't end up wasting all your time refactoring code you have and is working instead of writing new code.
Some of the reasons why one must refactor are:
Enhancing performance.
Organizing code.
You need to write your code in a different way to get something to work.
To do something in a different way because it saves a lot of work (i.e.: Using MXML instead of ActionScript).
You used the wrong name for a variable.
Consider learning some framework in whatever language you're using (or in any language for that matter).
I think that learning frameworks made my code a million times better. By learning the frameworks (and more importantly how they work) I learned not just design patterns, but how to implement them realistically.
Consider looking at rails, cakephp, or django (assuming you're in a scripting language; I don't know any desktop language frameworks. Sorry!). Then see how their pieces fit together.
I've heard a ton of complaining over the years about inherited projects that us developers have to work with. The WTF site has tons of examples of code that make me actually mutter under my breath "WTF?"
But have any of you actually been presented with code that made you go, "Holy crap this was well thought out!" or "Wow, I never thought of that!"
What inherited code have you had to work with that made you smile and why?
Long ago, I was responsible for the Turbo C/C++ run-time library. Tanj Bennett wrote the original 80x87 floating point emulator in 16-bit assembler. I hadn't looked closely at Tanj's code since it worked well and didn't require attention. But we were making the move to 32-bits and the task fell to me to stretch the emulator.
If programming could ever be said to have something in common with art this was it.
Tanj's core math functions managed to keep an 80-bit floating point temporary result in five 16-bit registers without having to save and restore them from memory. X86 assembly programmers will understand just what an accomplishment this was. Register space was scarce and keeping five registers as your temp while simultaneously doing complex math was a beautiful site to behold.
If it was only a matter of clever coding that would have been enough to qualify it as art but it was more than that. Tanj had carefully picked the underlying math algorithms that would be most suitable for keeping the temp in registers. The result was a blazing-fast floating point emulator which was an important selling point for many of our customers.
By the time the 386 came along most people who cared about floating-point performance weren't using an emulator but we had to support Intel's 386SX so the emulator needed an overhaul. I rewrote the instruction-decode logic and exception handling but left the core math functions completely untouched.
In my first job, I was amazed to discover a "safe ID" class in the codebase (c++), which was wrapping numerical IDs in a class templated with an empty tag class, that ensured that the compiler would complain if you tried for example to compare or assign a UserId into an OrderId.
Not only did I made sure that I had an equivalent Id class in all subsequent codebases I would be using, but it actually opened my eyes on what the compiler could do to guarantee correctness and help writing stronger code.
The code that impresses me the most, and which I try to emulate - is code that seems too simple and easy to understand.
It is damn difficult to write that kind of code. :-)
I have a funny story to tell here.
I was working on this Javaish application, filled with getters & setters that did nothing but get or set and interfaces and everything ever invented to make code unreadable. One day I stumbled upon some code which seemed very well crafted -- it was basically an algorithm implementation that looked very elegant = few lines of readable code, even though it respected every possible rule the project had to adhere to (it was checkstyled automatically).
I couldn't figure out who on the team could have written such code. I was dying to discuss with him and share thoughts. Thankfully, we had switched to subversion (from cvs) a few months earlier and I quickly ran am 'svn blame'. I loled all over the place, seeing my name next to the implementation.
I had heard stories about people not remembering code they wrote 6 months back, code that is a nightmare to maintain. I could not believe such a thing could happen: how can you forget code you wrote? Well, now I'm convinced it can happen. Thankfully the code was alright and easy to extend, so I've only experienced half of the story.
Some VB6 code by another programmer at my company I came across that handled the error conditions very well (whether it be deal with them directly or log them).
Along with some rather complex code that was well commented.
I know this will bring a lot of answers like,
"I've never find good code before I step in" and variations.
I think the real problem there is not that there isn't good coders or excellent projects out there, is that there's an excess of NIH syndrome and the fact that no body likes code from others. The latter is just because you have to make an intellectual effort to understand it, a much bigger effort than you need to understand you own code so that you dislike it (it's making you think and work after all).
Personally I can remember (as everyone I guess) some cases of really bad code but also I remember some pretty well documented, elegant code.
Currently, the project that most impressed me was a very potent, Dynamic Workflow Engine, not only by the simplicity but also for the way it is coded. I can remember some very clever snippets here and there, as well as a beautiful metaprogramming library based on a full IDL developed by some friends of mine (Aspl.es)
I inherited a large bunch of code that was SO well written I actually spent the $40 online to find the guy, I went to his house and thanked him.
I think Rocky Lhotka should get the credit, but I had to touch a CSLA.NET application recently {in my private practice on the side} and I was very impressed with the orderliness of the code. The app worked extremely well, but the client needed a few extensions. The original author had died tragically, and the new guy was unsophisticated. He didn't understand CSLA.NET's business object based approach, and he wanted to do it all over again in cut-and-dried VB.NET, without any fancy framework.
So I got the call. Looking at a working example of WinForm binding and CSLA.NET was pretty instructive about a lot of things.
Symbian OS - the old core bit of it anyway, the bit that dated back to the Psion days or those who even today keep that spirit alive.
And sitting right along side it and all over it is all the new crap created by the lowest bidders hired by the big phone corporations. It was startling, you could actually feel in your bones whether a bit of the code-base was old or new somehow.
I remember when I wrote my bachelor thesis on type inference, my Pascal-to-Pascal 'compiler' was an extension of a Parser my supervisor programmed (in Java). It had a pretty good structure as far as I can remember, and for me who had never done any serious Object-oriented programming, it was quite a revelation.
I've been doing a lot of Eclipse plug-in development and often had to debug into the actual Eclipse source code. While I haven't "inherited" it in the sense that I'm not continuing work on it, I've always been impressed with the design and quality of the early core.
I'm occasionally unfortunate enough to have to make alterations to very old, poorly not documented and poorly not designed code.
It often takes a long time to make a simple change because there is not much structure to the existing code and I really have to read a lot of code before I have a feel for where things would be.
What I think would help a lot in cases like this is a tool that would allow one to visualise an overview of the code, and then maybe even drill down for more detail. I suspect such a tool would be very hard to get right, given that is trying to find structure where there is little or none.
I guess this is not really a question, but rather a musing. I should make it into a question - What do others do to assist in getting their head around other peoples code, the good and the bad?
Hmm, this is a hard one, so much to say so little time ...
1) If you can run the code it makes life soooo much easier, breakpoints (especially conditional) break points are you friend.
2) A purists' approach would be to write a few unit tests, for known functionality, then refactor to improve code and understanding, then re-test. If things break, then create more unit tests - repeat until bored/old/moved to new project
3) ReSharper is good at showing where things are being used, what's calling a method for instance, it's static but a good start, and it helps with refactoring.
4) Many .net events are coded as public, and events can be a pain to debug at the best of times. Recode them to be private and use a property with add/remove. You can then use break point to see what is listening on an event.
BTW - I'm playing in the .Net space, and would love a tool to help do this kind of stuff, like Joel does anyone out there know of a good dynamic code reviewing tool?
I have been asked to take ownership of some NASTY code in the past - both work and "play".
Most of the amateurs I took over code for had just sort of evolved the code to do what they needed over several iterations. It was always a giant incestuous mess of library A calling B, calling back into A, calling C, calling B, etc. A lot of the time they'd use threads and not a critical section was to be seen.
I found the best/only way to get a handle on the code was start at the OS entry point [main()] and build my own call stack diagram showing the call tree. You don't really need to build a full tree at the outset. Just trace through the section(s) you're working on at each stage and you'll get a good enough handle on things to be able to run with it.
To top it all off, use the biggest slice of dead tree you can find and a pen. Laying it all out in front of you so you don't have to jump back and forward on screens or pages makes life so much simpler.
EDIT: There's a lot of talk about coding standards... they will just make poor code look consistent with good code (and usually be harder to spot). Coding standards don't always make maintaining code easier.
I do this on a regular basis. And have developed some tools and tricks.
Try to get a general overview (object diagram or other).
Document your findings.
Test your assumptions (especially for vague code).
The problem with this is that on most companies you are appreciated by result. That's why some programmers write poor code fast and move on to a different project. So you are left with the garbage, and your boss compares your sluggish progress with the quick and dirtu guy. (Luckily my current employer is different).
I generally use UML sequence diagrams of various key ways that the component is used. I don't know of any tools that can generate them automatically, but many UML tools such as BoUML and EA Sparx can create classes/operations from source code which saves some typing.
The definitive text on this situation is Michael Feathers' Working Effectively with Legacy Code. As S. Lott says get some unit tests in to establish behaviour of the lagacy code. Once you have those in you can begin to refactor. There seems to be a sample chapter available on the Object Mentor website.
I strongly recommend BOUML. It's a free UML modelling tool, which:
is extremely fast (fastest UML tool ever created, check out benchmarks),
has rock solid C++ import support,
has great SVG export support, which is important, because viewing large graphs in vector format, which scales fast in e.g. Firefox, is very convenient (you can quickly switch between "birds eye" view and class detail view),
is full featured, intensively developed (look at development history, it's hard to believe that so fast progress is possible).
So: import your code into BOUML and view it there, or export to SVG and view it in Firefox.
See Unit Testing Legacy ASP.NET Webforms Applications for advice on getting a grip on legacy apps via unit testing.
There are many similar questions and answers. Here's the search https://stackoverflow.com/search?q=unit+test+legacy
The point is that getting your head around legacy is probably easiest if you are writing unit tests for that legacy.
I haven't had great luck with tools to automate the review of poorly documented/executed code, cause a confusing/badly designed program generally translates to a less than useful model. It's not exciting or immediately rewarding, but I've had the best results with picking a spot and following the program execution line by line, documenting and adding comments as I go, and refactoring where applicable.
a good IDE (EMACS or Eclipse) could help in many cases. Also on a UNIX-platform, there are some tools for crossreferencing (etags, ctags) or checking (lint) or gcc with many many warning options turned on.
First, before trying to comprehend a function/method, i would refactor it a bit to fit your coding conventions (spaces, braces, indentation) and remove most of the comments if they seem to be wrong.
Then I would refactor and comment the parts you understood, and try to find/grep those parts over the whole source tree and refactor them there also.
Over the time, you get a nicer code, you like to work with.
I personally do a lot of drawing of diagrams, and figuring out the bones of the structure.
The fad de jour (and possibly quite rightly) has got me writing unit tests to test my assertions, and build up a safety net for changes I make to the system.
Once I get to a point where I'm comfortable enought knowing what the system does, I'll take a stab at fixing bugs in the sanest way possible, and hope my safety nets neared completion.
That's just me, however. ;)
i have actuaally been using the refactoring features of ReSharper to help m get a handle on a bunch of projects that i inherited recently. So, to figure out another programmer's very poorly structured, undocumented code, i actually start by refactoring it.
Cleaning up the code, renaming methods, classes and namespaces properly, extracting methods are all structural changes that can shed light on what a piece of code is supposed to do. It might sound counterintuitive to refactor code that you don't "know" but trut me, ReSharper really allows you to do this. Take for example the issue of red herring dead code. You see a method in a class or perhaps a strangely named variable. You can start by trying to lookup usages or, ungh, do a text search, but ReSharper will actually detect dead code and color it gray. As soon as you open a file you see in gray and with scroll bar flags what would have in the past been confusing red herrings.
There are dozens of other tricks and probably a number of other tools that can do similar things but i am a ReSharper junky.
Cheers.
Get to know the software intimately from a user's point of view. A lot can be learnt about the underlying structure by studying and interacting with the user interface(s).
Printouts
Whiteboards
Lots of notepaper
Lots of Starbucks
Being able to scribble all over the poor thing is the most useful method for me. Usually I turn up a lot of "huh, that's funny..." while trying to make basic code structure diagrams that turns out to be more useful than the diagrams themselves in the end. Automated tools are probably more helpful than I give them credit for, but the value of finding those funny bits exceeds the value of rapidly generated diagrams for me.
For diagrams, I look for mostly where the data is going. Where does it come in, where does it end up, and what does it go through on the way. Generally what happens to the data seems to give a good impression of the overall layout, and some bones to come back to if I'm rewriting.
When I'm working on legacy code, I don't attempt to understand the entire system. That would result in complexity overload and subsequent brain explosion.
Rather, I take one single feature of the system and try to understand completely how it works, from end to end. I will generally debug into the code, starting from the point in the UI code where I can find the specific functionality (since this is usually the only thing I'll be able to find at first). Then I will perform some action in the GUI, and drill down in the code all the way down into the database and then back up. This usually results in a complete understanding of at least one feature of the system, and sometimes gives insight into other parts of the system as well.
Once I understand what functions are being called and what stored procedures, tables, and views are involved, I then do a search through the code to find out what other parts of the application rely on these same functions/procs. This is how I find out if a change I'm going to make will break anything else in the system.
It can also sometimes be useful to attempt to make diagrams of the database and/or code structure, but sometimes it's just so bad or so insanely complex that it's better to ignore the system as a whole and just focus on the part that you need to change.
My big problem is that I (currently) have very large systems to understand in a fairly short space of time (I pity contract developers on this point) and don't have a lot of experience doing this (having previously been fortunate enough to be the one designing from the ground up.)
One method I use is to try to understand the meaning of the naming of variables, methods, classes, etc. This is useful because it (hopefully increasingly) embeds a high-level view of a train of thought from an atomic level.
I say this because typically developers will name their elements (with what they believe are) meaningfully and providing insight into their intended function. This is flawed, admittedly, if the developer has a defective understanding of their program, the terminology or (often the case, imho) is trying to sound clever. How many developers have seen keywords or class names and only then looked up the term in the dictionary, for the first time?
It's all about the standards and coding rules your company is using.
if everyone codes in different style, then it's hard to maintain other programmer code and etc, if you decide what standard you'll use have some rules, everything will be fine :) Note: that you don't have to make a lot of rules, because people should have possibility to code in style they like, otherwise you can be very surprised.
Closed. This question is opinion-based. It is not currently accepting answers.
Want to improve this question? Update the question so it can be answered with facts and citations by editing this post.
Closed 10 months ago.
Improve this question
I've never written functional specs, I prefer to jump into the code and design things as I go. So far its worked fine, but for a recent personal project I'm writing out some specs which describe all the features of the product, and how it should 'work' without going into details of how it will be implemented, and I'm finding it very valuable.
What are your thoughts, do you write specs or do you just start coding and plan as you go, and which practice is better?
If you're driving from your home to the nearest grocery store, you probably don't need a map. But...
If you're driving to a place you've never been before in another state, you probably do.
If you're driving around at random for the fun of driving, you probably don't need a map. But...
If you're trying to get somewhere in the most effective fashion (minimize distance, minimize time, make three specific stops along the way, etc.) you probably do.
If you're driving by yourself and can take as long as you like, stopping any time you see something interesting or to reconsider your destination or route, you may not need a map. But...
If you're driving as part of a convoy, and all need to make food and overnight lodging stops together, and need to arrive together, you probably do.
If you think I'm not talking about programming, you probably don't need a functional spec, story cards, narrative, CRCs, etc. But...
If you think I am, you might want to consider at least one of the above.
;-)
For someone who "jumps into the code" and "design[s] as they go", I would say writing anything including a functional spec is better than your current methods. A great deal of time and effort can be saved if you take the time to think it through and design it before you even start.
Requirements help define what you need to make.
Design helps define what you are planning on making.
User Documentation defines what you did make.
You'll find that most places will have some variation of these three documents. The functional spec can be lumped into the design document.
I'd recommend reading Rapid Development if you're not convinced. You truely can get work done faster if you take more time to plan and design.
Jumping "straight to code" for large software projects would almost surely lead to failure (as immediatley starting posing bricks to build a bridge would).
The guys at 37 Signals would say that is better to write a short document on paper than writing a complex spec. I'd say that this could be true for mocking up quickly new websites (where the design and the idea could lead better than a rigid schema), but not always acceptable in other real life situations.
Just think of the (legal, even) importance a spec document signed by your customer can have.
The morale probably is: be flexible, and plan with functional or technical specs as much as you need, according to your project's scenario.
For one-off hacks and small utilities, don't bother.
But if you're writing a serious, large application, and have demanding customers and has to run for a long time, it's a MUST. Read Joel's great articles on the subject - they're a good start.
I do it both ways, but I've learned something from Test Driven Development...
If you go into coding with a roadmap you will get to the end of the trip a helluva lot faster than you will if you just start walking down the road without having any idea of how it is going to fork in the middle.
You don't have to write down every detail of what every function is going to do, but define you basics so that way you know what you should get done to make everything work well together.
All that being said, I needed to write a series of exception handlers yesterday and I just dove right in without trying to architect it out at all. Maybe I should reread my own advice ;)
What a lot of people don't want to admit or realize is that software development is an engineering discipline. A lot can be learned as to how they approach things. Mapping out what your going to do in an application isn't necessarily vital on small projects as it is normally easier to quickly go back and fix your mistakes. You don't see how much time is wasted compared to writing down what the system is going to do first.
In reality in large projects its almost necessary to have road map of how the system works and what it does. Call it a Functional Spec if you will, but normally you have to have something that can show you why step b follows step a. We all think we can think it up on the fly (I am definitely guilty of this too), but in reality it causes us problems. Think back and ask yourself how many times you encountered something and said to yourself "Man I wish I would have thought of that earlier?" Or someone else see's what you've done, and showed you that you could have take 3 steps to accomplish a task where you took 10.
Putting it down on paper really forces you to think about what your going to do. Once it's on paper it's not a nebulous thought anymore and then you can look at it and evaluate if what you were thinking really makes sense. Changing a one page document is easier than changing 5000 lines of code.
If you are working in an XP (or similar) environment, you'll use stories to guide development along with lots of unit and hallway useability testing (I've drunk the Kool-Aid, I guess).
However, there is one area where a spec is absolutely required: when coordinating with an external team. I had a project with a large insurance company where we needed to have an agreement on certain program behaviors, some aspects of database design and a number of file layouts. Without the spec, I was wide open to a creative interpretation of what we had promised. These were good people - I trusted them and liked working with them. But still, without that spec it would have been a death march. With the spec, I could always point out where they had deviated from the agreed-to layout or where they were asking for additional custom work ($$!). If working with a semi-antagonistic relationship, the spec can save you from even worse: a lawsuit.
Oh yes, and I agree with Kieveli: "jumping right to code" is almost never a good idea.
I would say it totally "depends" on the type of problem. I tend to ask myself am I writing it for the sake of it or for the layers above you. I also had debated this and my personal experience says, you should since it keeps the project on track with the expectations (rather than going off course).
I like to decompose any non trivial problems loosely on paper first, rather than jumping in to code, for a number of reasons;
The stuff i write on paper doesn't have to compile or make any sense to a computer
I can work at arbitrary levels of abstraction on paper
I can add pictures and diagrams really easily
I can think through and debug a concept very quickly
If the problem I'm dealing with is likely to involve either a significant amount of time, or a number of other people, I'll write it up as an outline functional spec. If I'm being paid by someone else to develop the software, and there is any potential for ambiguity, I will add enough extra detail to remove this ambiguity. I also like to use this documentation as a starting point for developing automated test cases, once the software has been written.
Put another way, I write enough of a functional specification to properly understand the software I am writing myself, and to resolve any possibile ambiguities for anyone else involved.
I rarely feel the need for a functional spec. OTOH I always have the user responsible for the feature a phone call away, so I can always query them for functional requirements as I go.
To me a functional spec is more of a political tool than technical. I guess once you have a spec you can always blame the spec if you later discover problems with the implementation. But who to blame is really of no interest to me, the problem will still be there even if you find a scapegoat, better then to revisit the implementation and try to do it right.
It's virtually impossible to write a good spec, because you really don't know enough of either the problem or the tools or future changes in the environment to do it right.
Thus I think it's much more important to adapt an agile approach to development and dedicate enough resources and time to revisit and refactor as you go.
It's important not to write them: There's Nothing Functional about a Functional Spec