question

I'm not sure I'm following your description. Pasting in the utterance file would help. (Or a link to a pastebin of it.) In general, Alexa performs poorly when you have two slots in one utterance without any static text between them. That could be an issue.

Are you trying to use anything (invocation or value for one of the slots) that isn't either a real English word, or a very commonly used proper noun? The reason I ask is that from what we can tell, the way the speech matching works is not linear - it doesn't say "Step 1: invocation matches. Step 2: First slot matches. Step 3: final slot matches.". Instead, it seems to be calculating the entire utterance's compatibility (from "Alexa..." all the way to your last word) against everything you have enabled - your skill, other skills, built-in/first-party functionality. It's possible that the positives you're getting are just barely above the requisite confidence level cutoff, and the false negatives are just below it. If you had an ambiguously phonetic invocation name, this would often result in the utterance string in the unknown request card in your alexa app to show an invocation name that isn't yours. We had that problem quite a bit with InsultiBot - when it matched properly, the card would show it as something like "Ask InsultiBot for a compliment", but if it failed to match on a very similar voice input, the card might show something like "Ask in salt a bought for a couple mint". It wasn't that the voice input was THAT different, it was just that Alexa was determining that the user wasn't asking InsultiBot a question since it wasn't close enough, and therefore didn't show the response as being in any way tied to InsultiBot's utterances. We were able to greatly increase our hit rate by changing our invocation word and updating the sample values for our slots. If you paste your utterances and invocation name, as jjaquinta mentioned, it would be easier for us to help.