The trick is how much state you have to keep in your head. Something like Blender requires enormous state about even things like what key combinations mean in different contexts: it's a poster child for impossibly demanding state requirements.

Something like the Flash pen tool is much simpler, but still absolutely requires you to maintain some state: clicking versus click-dragging, remembering not to close a shape by simply clicking on a control point without also dragging out Bezier points. There are expectations before you can begin to flow with the thing.

I've been working hard on this concept using a Minecraft mod (Snowball Madness: http://www.airwindows.com/snowball-madness/ ) that suffered the same problem. I'd made countless 'effects' so it was nearly impossible to remember what did what.

After a drastic functionality-culling process, I began rebuilding things in line with a concept: generalizing. If you can place a block above a snowball and it places the block where the snowball hits, that's what it does, no exceptions. TNT used to spawn explosions just for silly fun, but it became 'place TNT block' altering the type of silliness. Pickaxes used to dig large holes in rock (in some cases, leaving ores hanging) so all the other tools got similar treatments: axes vanishing wood logs, shovels vanishing dirt, hoes turning grass/dirt into tilled farmland. Always trying to incorporate 'cheaty' ways of doing things but predictably so.

It's like the old Apple UI guidelines. The default expectation is that you can grope blindly towards a result and things do what you think they would do, allowing you to not think about the process.