For me, the book was one I bet nobody else has heard of, the Engineering Mathematics Handbook of Jan J. Tuma. It is a reference work, with page after page of identities and formulas, organized by subject, and accompanied with hundreds of handsome two-color diagrams. It is aimed at the practicing engineer, who doesn't want to have to remember or derive the Taylor expansion of $\sinh x$; he wants to look it up. But that was not how I used it.
The first page starts with algebraic identities, such as the commutativity of addition, and then it works its way up through geometry, trigonometry, calculus, and differential equations, with excursions along the way into probability theory, statistics, and engineering subjects such as mechanics.

I spent hundreds of hours with this book, tinkering with the identities and seeing how they related. For example, there are two pages that tabulate the derivatives of common functions. One of these is $$\left(u^v\right)' = vu^{v-1}u' + u^vv'\ln v$$ and by substituting various choices of $u$ and $v$ you can see how the more familiar examples $$\begin{align}\left(a^x\right)' & = (\ln a)\cdot a^x\\
\left(x^n\right)' & = nx^{n-1}\end{align}$$
(which are also tabulated) are special cases of the general one.
Every entry was a puzzle to be solved. Page 85 contains a table of the derivatives of the elementary trigonometric functions, including a bald assertion that $(\sec u)' = -\frac{u' \sin u}{\cos^2 u}$. Well, I knew the definition of $\sec x = \frac{1}{\cos x}$; could I work out the correct derivative myself, using the chain rule, and would it be what the book said? Yes, I could! It was! But wait, someone else told me the derivative was $-\sec x\cdot \tan x$. Oh, but that is when $u=x$, and yes, it checks out.
There is a section with diagrams of geometric figures, and the positions of their centroids, their moments of inertia, and so on. This was preceded by a page that gave the general formula for the centroid. If I applied the definition and carried out the calculation of the centroid of a parabolic segment, would I get the answer that was written next to the picture of the parabola? Yes! Could I also calculate the moment of inertia? No! (I never did figure it out, but I tried hard.) Another page gave the classification of quadratic curves by shape. If I put the coefficients of $xy-1=0$ or $x^2+y^2=25$ into the formulas, do they correctly identify the curves as a hyperbola and a circle? (They did!) Was the classification invariant under translations and rotations? (It was!) How did it work? (Parts were clear, other parts less so.) How does it treat degenerate cases, such as $x^2+y^2 = -25$? This is how I learned that quadratic curves have a hidden complex part.
Some pages were just mysteries. The definition of a hermitian matrix was presented without comment. The meaning was clear, but the purpose was not. I didn't find out why Hermitian matrices were interesting until thirty years later. There were tables of Laplace transforms. What were they for? But they were mysteries that exercised my mind, and when I encountered the answers, I smiled and said “Oh, so that's what that is about!”
The book would not have inspired many people, perhaps, but when I met it, it was the right thing at the right time, and I loved it then and still do.