The post constituting the OP is really broad and touches on a lot of stuff, suggesting that the question is about monolingual elicitation, or how do you start eliciting, or is there a formula that everybody uses. What I think is the intended question is something like "What does it take to create a grammar based on elicited data?". That's the matter that I will address.
The most difficult problem (also the last problem) is deciding what constitutes "good enough". I commend Cowell's grammar of Syrian Arabic, because the author decided to go the extra mile or so. I have seen plenty of lousy grammars (name will be withheld). I assume that the goal is to create a Cowell-quality (or, Whitney Sanskrit quality) grammar. The related problematic point is determining whether your goal is to say what the language is like, or to show what the language is like. Whitney is based on excellent knowledge of the language, but not a corresponding level of exemplification. Cowell is chock full of examples. So you have to decide what kind of a grammar you plan to produce. Newman's encyclopedic grammar of Hausa is kind of unique, and clearly can only be done for a language that is already well-described (contra the "undescribed" assumption).
Now for some numbers. I have approximately 400 hrs of elicited data on Logoori, covering about 120,000 items (not just sentences... basically, "lines" including words, semantic datoids, NPs and full sentences (or two). I have an extracted collection of 983 noun stems and 707 verb stems (probably another 200 items lurking in the database).
This was carried out over about 7 working years, and the data-elicitation part which covers phonology and morphology and rather elementary syntax is maybe 80% finished. This includes zero texts, though I have some published words (stories and bibles) that I've consulted. In light of the highly variable nature of the language, it would be preposterous to get "the complete picture" for "all speakers". The current grammar fragment is about 700 pages, and does not try to cover systematically syntax (extraction, complementation, conjunction, syntax of agreement, etc.).
The data is starkly divided into the written part of about 25,000 items, not recorded just directly transcribed into notebooks, from the first speaker, and the recorded and post-hoc transcribed part from everybody else (10 speakers). The reason for this is that technology improved over the years: the first chunk was conducted from 2005, and the latter chunk started in 2014. Recording technology has improved orders of magnitude, and I find it easier, now but not then, to record sessions and not waste face-time with transcribing.
One way to know the meaning of a word in the "record then transcribe" regime is to ask "How do say 'chicken'?" and transcribe what the guy says. This works if your voice is also picked up on the recordings: you are in trouble if your stimulus is not included (this does happen). Or, if the guy also gives the English translation, that gives you some evidence what the word means. You have to treat "meaning" as a hypothesis that you are continuously trying to refine. I am pretty happy with my understanding of the completive variants of verb-forms which I spent a lot of time on. This is one of those YMMV matters. These facts are both because of the nature of this particular language, and because of my goal and standards in writing a phonology-and-morphology of this language.
There are plenty of grammars that are less Cowell-like and more Whitney-like (or even shorter). Whether one calls the word a grammar, a sketch, an introduction, essentials or some other expression is up to the author.
