Most of the Haskell parsing libraries return the failing string if an error occurs. So you could pass this to the slower but more friendly parser to give you more information.

Obviously, this bloats the code hugely, but that may be acceptable in some circumstances.

You might even extend the technique so that you have a very fast parser that works for 95% of the code, and hand over to a slower parser for other cases.

I appreciated that Attoparsec had straightforward Applicative and Alternative instances; many parsers have ended up with specialized versions of the operators or whole new shadow classes (I believe the Utrecht parser combinators have their own Applicative class).