Parser combinators are a functional abstraction for parsing that abstracts hand-written recursive-descent parsers behind a high-level set of combinators. While these kinds of parsers are popular in the functional programming community, they have been historically criticised: * Parser combinator performance is sub-par compared with handwritten parsers. * The high-level grammar is obscured by the combinators compared with parser generators. * The error messages generated by parser combinators are not of fantastic quality. This dissertation addresses each of these complaints with the work split across two libraries, both called `parsley`: one in Haskell and the other in Scala. Within these libraries, different issues are tackled. Haskell `parsley` addresses the performance concerns of parser combinators by modeling them as a strongly-typed deep embedding, allowing for optimisations and analysis to be performed. To eliminate the overheads of interpretation and achieve high performance, `parsley` makes use of staged metaprogramming to convert the continuation-passing style automaton into Haskell code resembling hand-written recursive descent parsers; this is faster than contemporary parser combinator libraries in Haskell. Writing parsers is often an ad-hoc exercise; this dissertation introduces parser combinator design patterns that help structure and standardise how these parsers should be written. These patterns focus on a handful of issues: cleanly handling precedence hierarchies and expression parsing; organising and distinguishing between low-level tokens and higher-level parsing; and abstracting away semantic actions and meta-data processing. This helps to make clean, maintainable, parsers. Finally, Scala `parsley` has focused on improving the design of parser combinator error systems. The design of this improved system is explored as well as how it can be implemented efficiently, minimising impact on the "happy path." This gives rise to more parsing patterns as well as enriching existing patterns to incorporate errors. The new patterns help provide the tools to build bespoke, descriptive, errors.