diff --git a/notes.md b/notes.md
index 8319fc0..2e22d3a 100644
--- a/notes.md
+++ b/notes.md
@@ -11,3 +11,28 @@ functions. I can think of these off the top of my head:
 * Assembler loop - takes structure above and outputs one word
     - Maybe should wrap in a sentence loop, line-by-line loop, paragraph loop,
       etc.
+
+Syntax aware generate is actually pretty bad. I think it forces it to be too
+random. The POS tagging is too error prone and fine-detailed.
+
+Ideas for the future:
+
+Pick one or two lines of the haiku from actual haiku or other poems. Then add a
+line or two from the corpus (e.g. trump tweets) that both fits the syllables and
+rhymes with the end(s) of the real poetic line. I think both sources could be
+ngram generated, but I think it would be ideal if they were picked wholesale
+from the source. The problem with that approach is that you'd also have to find
+a common word between the two source extractions so that the sentence doesn't
+abruptly shift between lines. Or, maybe that's a good thing? I guess I should
+try both.
+
+Maybe try just switching out the nouns, verbs, adjectives, and adverbs leaving
+the rest of the sentence structure largely intact after the tree replace?
+
+Use word similarity vectors to construct a sentence (or poem) around a central
+theme. E.g. construct something like the buffalo sentence: syntactically correct
+sentences, and maybe even semantically meaningful sentences, but in a totally
+novel form because of some arbitrary restriction (can only use the word
+"buffalo", or animal words, or onomatopoeias, or etc.).
+
+Integrate alliteration and rhyming.