You have a mnemonic that is n-bits (eg. 128-bits for 12 words) take those bits and encrypt them with AES256 with a key derived from a strong passphrase and a salt derived from the address (like BIP38).
Now you get 128-bits of encrypted data (encryption without IV) which you can encode to 12 words using the same BIP39 scheme. All you have to do is write down these words as if they were your mnemonic.
To import to a wallet you just decrypt these words and convert to an un-encrypted BIP39. That is decode 12 words to get the byte[], derive the AES key from the passphrase and address like above, decrypt using AES256. Now you have un-encrypted 128-bit entropy. Encode this using BIP39 scheme and you have the original words.
that's a pretty neat idea the only potential problem I see with it is, it's not very "portable". It requires a fixed wordlist. what if there is no wordlist at all? or what if the wordlist is different than the bip39 one. that means we have to store a complete copy of whatever wordlist it is we want to use and somehow the software has to be smart enough to know what wordlist we want it to use?
i'd much prefer a solution that
1)didn't require a fixed wordlist
2)allowed for variable length mnemonics
3) but still operated by changing words in to other words