diff --git a/notes.md b/notes.md
index 8319fc0..411b249 100644
--- a/notes.md
+++ b/notes.md
@@ -11,3 +11,20 @@ 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?