The standard advice is to make sure you understand something before you make a card about it, but I've never seen an adequate description of what people mean by "understand" in this context. If a card fails (but follows the other rules), it must mean that you didn't properly understand in the first place, which isn't particularly actionable.
There are two distinct kinds of knowledge: tacit knowledge, the ability to perform some task, and rote knowledge, the ability to recall facts. In most domains, tacit knowledge more useful and harder to come by than rote knowledge. I have come to believe that "understanding" in the SRS context means learning how to do something, how to select an appropriate fire extinguisher in your example. Cards formulated according to those guidelines, then, encode related facts. The idea seems to be that recalling the facts will also reinforce the tacit memory.
An alternative is reinforcing the tacit knowledge directly, which requires a completely different approach: Each card should be a prompt to perform some task you care about, with all relevant context (and possibly extraneous context). This has a reasonable chance of succeeding if you use a premade deck, because all of the information is present. Unfortunately, this isn't how most shared decks are structured.
I've been using a variant on incremental reading for this: I took some books from Project Gutenberg and split the text up into cards and blanked out a word towards the each passage as a Cloze deletion. If I'm successfully following along with the argument that the author is making, then I can usually predict the missing word despite having never read the passage before; otherwise, I need to spend more time with the passage.
Other skills have separate decks: I have a piano deck, for instance, that contains sheet music on the front of the cards with a few bars marked for special attention (or blacked out). I review that one on a tablet while actually at the piano, playing the pieces. I use a photo editor to split up the scan into small parts and throw them all into Anki; it seems to work just fine without the pre-learning that's usually recommended.