Recently, I became interested in trying to document the grammar and phonologies of colloquial or "street" forms of English. Is there an easy way to figure out how people in my neighborhood actually talk (as opposed to how I talk, or how grammar textbooks talk). People will probably speak differently than they would naturally if they notice I'm taking notes on how they talk on a clipboard! So what would be a good way to observe it easily? It's hard to pin down colloquial forms of languages.
No, there is no easy way to make a reference grammar of a "colloquial" language, in fact I think that it is impossible to do so, but you could dial back your aspirations and write "a good descriptive grammar". Reference grammars have an elevated air of definitiveness, authority and officiality which contradicts the nature of a "colloquial" language form.
The lowest level of grammatical description would be some kind of tagged and commented fixed corpus of the language. Basically, you record speech and tag all of your examples for whatever the relevant feature is observed in the utterance. You will have to somehow figure out what things need to be mentioned. If you can say "might should" in your dialect and the people you're recording do so too, you may not realize that the construction is a bit colloquial and regional, and you could miss some interesting details. There are tens of thousands of things to be on the lookout for – you might should consult with a dialect specialist, if there are any.
There are a few general precautions that need to be taken in order to get good data (or you have bad data, you'll have a bad description). One thing is to be sure that the people you're recording do speak the dialect, or are speaking the dialect. Also, you want to avoid influencing their speech by interjecting your fancy upper class dialect (if that's what you speak). One way that people have addressed the observer-contamination problem is to make passive recordings of "natural speech", for example recording conversations at a barber shop (not a hair salon). Another related method is to engage speakers in long discussions, especially of subjects that they will have strong feelings about. The underlying reasoning here is that you want them to forget that they are being recorded, so that they talk more naturally. There is a wealth of reading to do about sociolinguistic interview techniques.
Another detail is that a good account of a colloquial language form will be accurate with respect to phonology. You cannot just write "cat" when the speaker says the word "cat", you have to learn how to transcribe it phonetically – narrow transcriptions, not simply "phonemic transcriptions" because you don't actually know what the phonemes and allophones are until you have a narrowly-transcribed corpus and an analysis that tells you the predictable redundancies in the corpus.
A description based on 5,000 utterances will not be particularly good though it will be somewhat informative. Increasing to 1,000,000 utterances would be much better, but sort of too big to do in a lifetime. One of the problems with a small corpus is distinguishing "not in the language" from "I didn't observe it". In regular field work situations working with a speaker of a foreign language, you simply ask "Can you say [xyz wrtq] in your language?". When dealing with "colloquial" forms of language (non-standard varieties not taught in schools or given any particular respect), asking "Can you say 'We couldn't might shoulda go'?" may just annoy the person you're interviewing, because they will think that you are mocking them.
IMO a better approach is to observe the dialect from a bit of a distance, then identify a small and well-defined area of the language to look at. Pronunciation is the easiest because it is literally what all of your primary data will be, but also the hardest because you have to learn some rather arcane skills for transcribing and even doing acoustic analysis, if you are going to make any generalizations about "what goes on in this dialect".
Perhaps (as already hinted) by compiling a reasonably representative corpus of the variety you are after. If you want just the grammar, you do not even need a phote[tm]ic transcription, but an orthographic one is enough (and even better in some respects).
One important rule when doing the recordings (since you want recordings, to be transcribed later) - never ever mention you are making a linguistic research. If you have to, and do not want to lie, tell the respondents you are making a sociological research, or something similar (after all, that would be one of the uses of such a corpus). Engage in a dialogue, and try to make them speak freely. Politics is a good topic, but can be explosive and dangerous sometimes. Steering the topic towards urban development is safer. Recalling good ol' days works best (with elderly people). Be creative. Of course, do not forget to ensure everything is legal (i.e. consent, preferably written, with the recording taking place)
Such an orthographic transcription (done manually) usually takes several times the recording wall time, so scale up your research team (heh) accordingly (40 times for a reasonably broad phonemic transcription - personal experience, and the language was easier to transcribe than English). It would be much easier these days with ASR systems exceeding human accuracy, though (but then depends on the difference from the "standard" language).
Overall, you are looking into several person-years to even compile a reasonably sized corpus, good for the most basic research only, alas.