He's opposed to BAD functional specs, which is fine, but too many people read his title and decide it's OK to skip to writing the code first.
The theory there is that written documents will always be out of date, incomplete, stale, and open to multiple interpretations, so we might as well just accept that the face-to-face communication is infinitely higher bandwidth and will have to happen anyway and not waste the time trying to write it all down up front.
So if the PM needs to spec for their own edification, that's fine: do whatever works for you. But I, as the developer, will trust the PM themselves as the source of truth as to what the feature should do, and it's the PM's job to know what they want well enough to answer my questions real-time as I'm building.
On the other hand I've also had the unhappy experience of wading through dense, arcane specification documents written in stilted language ("the system shall...." ad nauseum) that are ultimately not much better.
In the end I think what this illustrates is that writing a good, concise, clear, complete functional specification is not easy. It requires a talented PM and no formulaic approach, whether it is agile or anything else, can get around that.
It definitely still requires good product managers, but again the idea is that no matter how good the spec or whatever is you'll still always have questions, and it's better to just build that fact into the process and get people used to asking questions rather than having people make assumptions. And the best way I've seen to make sure that the PM has the answer to those questions is the estimation session, since if the PM doesn't know what they want well enough to answer the questions quickly as you develop, you'll find that out in the estimation session before anyone is tasked with working on the feature. That at least largely avoids the situation where a developer starts working on something and finds out the PM really didn't have a great idea of what they wanted after all, or that it's technically infeasible, which (in my experience) is just a really bad situation to be in.
The story approach still presents endless ways to screw up, but it's far better than any other approach we've tried over the last 7 or so years.
A complete, unambiguous, and functional specification is called a program. May or may not be "good".
The going back and forth that you describe as inefficient is where the collaborative creative act occurs, where you work with the person that needs a custom system to define what you can build for them.
The functional spec usually comes before the html code.
Sure, Fried treats his html code as a functional spec, but others who work differently would not be able to find what they are used to calling a functional spec.
Spolsky: "feature"
Fried: "app"
Suppose either is designing GreenfieldGloryApp - he might start with a one-page vision statement. He might then make some very high level mock-ups of scenarios. Maybe they would even be done in HTML. This is where they diverge: At this point 37signals would start building. Fog Creek would iterate 'specs' many times further by drilling down to feature or feature-set level functional specs. Each feature-level spec would start with a vision statement and then be fleshed out with mocks and "detailed descriptions".
-----------
Joel Step 1:
You start at the highest level: a vision statement, no more than one page explaining the gist of the new feature.
Fried Step 1:
So what do we do in place of a functional spec? We write a one page story about what the app should do. If it takes more than a page to explain it, then it’s too complex.
-----------
Joel Step 2:
Once that’s nailed down, you can develop storyboards… mockups of the screens showing the user’s progression through the application, with detailed notes showing how they work.
Fried Step 2:
Then we begin building the interface — the interface is the functional spec. First with some quick and simple paper sketches, then directly into HTML. Unlike paragraphs of text that are open to alternate interpretations, interface designs are common ground.
-----------