Since you mentioned you were unsure how to do sentences AJATT style, I'll tell you how I do them. It's not necessarily the best way for you, but it's a starting point.
Stage 1: Basic vocabulary/grammar. I used vocabulary lists to find about a thousand basic words and simply did the traditional "Spanish class" method of having English on one side, Japanese on the other (I did these cards both ways, J->E and E->J). More specifically... at this stage, I clumped both the kanji and the hiragana both on the J side, even on J->E cards. I even used romaji sometimes, but I highly disrecommend that now. I also made grammar flashcards, eg. Q: "Past tense of a ぶ verb" A: "ぶ --> んだ"
Stage 2: Basic sentences. I started making sentence flashcards, mostly using sentences from Tae Kim's Guide (
http://www.guidetojapanese.org/). The question side would contain the sentence, the answer side would contain readings of the words, and a translation. When a sentence contained new words (which was almost always), I'd add another J-E/E-J pair for that word.
Stage 3: Intermediate sentences and clean-up. In this stage sentences are added in bulk (avg. at least 25 a day), mostly from Yahoo辞書. Grammar cards from Stage 1 are deleted as they show up. E->J cards are from stages 1 and 2 are deleted as they show up. With J->E cards, the reading is moved from the Question side to the Answer side; the English translation is turned into invisible text (see appendix below). In a few cases, I replaced the English translation with a picture or row of pictures, but I stopped this as it was just sooo much effort. Old stage 2 sentences are modified so the translation is invisible (see appendix below). With new words in a sentence, rather than make a J-E card for that word with no context, I'd search for more sentences with that word. Eg. a new sentence has the word 女王, which I don't know, then I'll add some other sentences that also have 女王.
Stage 4: Advanced sentences, contexts, romaji dictation (experimental). In this stage, sentences are added not just from Yahoo辞書 and Tae Kim, but also Space ALC, and "real" Japanese (in my case, old SNES videogames like Chrono Trigger). For the "real" sentences, instead of putting an invisible translation, I put an invisible "context".
Example: Q: ああ。ケフカを倒し、平和な世界を取り戻す!!
A: <(Sabin, after Celes rescues the kid in Tzen)>倒す 【たおす】
平和 【へいわ】
取り戻す 【とりもどす】
Notice how in the example, I make no attempt at an English translation, but rather a context. Romaji dictation: I'm experimenting with this, it's a variation on the audio dictation. Basically, the "Q" side of the card has a sentence in romaji, with no spaces, the "は" particle is written "wa", the "を" particle is written "o", the "へ" particle is written "e", and the "answer" is the sentence in Japanese, and you try and write it based on just the "romaji soup".
Another technique I invented in this stage is what I call WWWJDIC mining. It's known that the actual Tanaka Corpus is flawed so I don't use that, however, if you just enter a word in kanji, you'll get lots of actual entries using that word, and the entries themselves at least require peer review (just don't add ones which say "User submitted, awaiting review"). This WWWJDIC mining lets you get lots of clauses involving your word extremely fast. For example if you're adding 情報, search for it in WWWJDIC and get some clauses like...
安全情報 【あんぜんじょうほう】 (n) safety information; safety bulletin
遺伝情報 【いでんじょうほう】 (n) genetic information
会社情報 【かいしゃじょうほう】 (n) corporate information
各情報 【かくじょうほう】 (n) all information
These are a really excellent supplement to the sentences, they're short and numerous and really good for learning readings, they'll also help you learn the evil elusive 連濁.
EDIT: By the time you reach stage 4, you'll be finding you "know" new words as you encounter them even if you never saw them before. Because they're made up of kanji you've seen in a dozen different words. This is when the kanji start to become really nice and you say, "Gee I'm glad Japanese uses kanji instead of an icky, primitive *alphabet*!"
APPENDIX
Using invisible text. I dunno how or if this works on other SRS's besides Mnemosyne. But it should work on any SRS that parses html. It's a neat trick I discovered, to make invisible text, just enclose it in <" ">. Mnemosyne will parse it as html, but it's meaningless so it becomes blank. You only see it when you click "edit card". So it lets you make translations and notes invisible unless you actually need them. Very useful for entering the "Japanese Trance" where you see no English and think no English for long stretches at a time.