Natural selection is the observation that 'nature' can work like an animal or plant breeder to modify traits - and the suggestion that this process could (given enough generations) lead not just to amplified, suppressed or re-shaped traits, but to new forms (such as new species or families of organisms).
The strength of natural selection is its accumulative potential: natural selection builds incremental change upon incremental change so that 'mony a mickle maks a muckle' (lots of littles makes a lot).
An analogy would be an animal trainer that begins with undirected behavioural variations and, incrementally, by small but many steps, uses reinforcement to build long and complicated sequences of behaviour, as when a squirrel is trained to do an obstacle course:
So, more-or-less-plausible stepwise incremental and accumulative sequences of natural selection have been constructed to explain the evolution of something as complex as the mammalian eye.
The constraints of natural selection include that what it builds-upon must be un-directed - so that the original form/s of which all later form/s are incremental shapings must have arisen by chance - specifically the original form must have arisen with no aim or teleology.
So that the first step in the presumed evolution of life on earth must have happened without natural selection - on the one hand, it must have been a form that was sufficiently simple to have happened by undirected chance arrangements of matter, yet on the other hand that which arose by undirected chance must have had a form sufficiently complex to be capable of replication.
(This whole problem is very well described, and a solution proposed, in my old friend Graham Cairns-Smith's book 'Seven Clues to the Origin of Life'.)
The other major constraint of natural selection (which does not apply to artificial selection, i.e. teleological selection done by an intelligence such as a human) is that each modification must provide an improvement in differential reproductive success - each incremental step must enhance relative replication.
Whether or not natural selection is a possible explanation for any given change under discussion is ultimately a matter of probabilities - except that the data for calculating probabilities is almost never available, and perhaps is not an objective fact but itself an assumption.
And even when plausible probabilities have been calculated - for instance, by GC-S in the above-mentioned book; which purports to show that the energetic properties of nucleic acid and amino acid/ protein molecules mean that neither could be the first step in the evolution of life - that this is too wildly improbable to have happened by chance during the believed-lifespan of the earth ... then nobody seems to take any notice of the results.
The fact is that the constraints of natural selection are not interpreted, in practice, as placing constraints on its applicability - since within the field of biology natural selection and physical explanations such as chance or the intrinsic properties of stuff (maths, physics, chemistry, systems etc), are the only two allowed classes of explanations for anything.
No matter how implausible or improbable, the assumption is that either undirected chance alone, or undirected chance plus natural selection, must suffice to explain everything.
So the constraints on natural selection are not interpreted as constraining its applicability as a general phenomenon - since NS is assumed to be generically universally sufficient to explain everything; but rather as a constraint upon the plausibility of any specific theory within natural selection or physical explanations.
(Apparent exceptions to this statement prove, on closer examination to be sub-classes of either or both natural selection/ physical theories - or else incoherent nonsense.)