Say you're in a meeting talking about the tech for your game, and you propose doing a new realtime radiosity lighting engine (for example). Someone else says "too risky, it will take too much time, etc. let's go with a technique we know".
So you go home and in your night hours you code up a prototype. The next day you find a boss and go "look at this!". You show off some sexy realtime radiosity demo and propose now it should be used since "it's already done".
Dirty, dirty coder. The problem is that it's very hard for the other person to win the argument that it's too hard to code once you have a prototype working. But in reality a sophisticated software engineer should know that a demo prototype proves nothing about whether the code is a sensible thing to put into a big software product like a game. It's maybe 1% of the total time you will spend on that feature, and it doesn't even prove that it works, because it can easily work in the isolation of a demo but not be feasible in real usage.
I used to pull this trick all the time, and usually got away with it. I would be in a meeting with my lead and the CTO, the lead would not want some crazy feature I was pushing, and then I would spring it on him that it was "already working" and I'd show a demo to the CTO and there you go, the feature is in, and the lead is quietly smoldering with futile anger. (the most dickish possible coder move (and my favorite in my youth) is to spring the "it's already done, let me show you" as a surprise in the green light meeting).
But I also realize now that my leads had a pretty good counter-play to my dirty trick. They would make me put the risky feature they didn't like in an optional DLL or something like that so it wasn't a crucial part of the normal production pipeline, that way my foolishness didn't impact lots of other people, and the magnified workload mainly fell only on me. Furthermore, optional bits tend to decay and fall off like frostbitten toes every time you have to change compiler or platform for directx version. In that way, the core codebase would clean itself of the unnecessary complicated features.