There should be a clear separation between code and other stuff. The most obvious thing to do is move all code to a src/ directory.

Some people might feel more motivated to build the project using Cabal rather than the crude Makefile. This would make it more of a standard Haskell package.

How do other Haskell projects achieve this? Should the library (or libraries) be packaged, or should there be a single package for the whole program?

The current way of doing analysis doesn't take into account the fact that natural language texts are composed of sentences. This is obviously not acceptable. The user should be able to generate individual sentences, and this should be reflected in the way analysis is done.

This issue is closely related to #3, as sentence analysis can't be done if chains are not built iteratively.

Can other methods be used in conjuction with Markov chains to generate text that has a closer resemblance to text generated by humans? Issue #5 addresses this by analyzing sentences instead of arbitrary chunks of texts.

Could Markov text generation be done based on parsing trees or some other NLP technique? How do we measure relative performance of one method versus another?

Create a libmonkey (or something similar) as a generic API for building Markov generators. This has the major advantage of making Writer Monkey useful for other applications than Markov text generation (e.g. for music, graphics, various forms of text etc.)

Right now, Monkey.Analysis builds Markov chains from a large chunk of text in one big pass. This works, but it isn't very flexible. The programmer should be able to enhance or strip the model however he/she pleases. This requires support for building of Markov chains "on the fly", as new text is added.

Ideally, this would be implemented as a monoid over Markov chains (if such a thing exists).