I'm pretty sure you are talking about Hermite reduction. The construction is not the same as Sylvester; we are not making any distinction between positive and negative diagonal elements, indeed a field cannot be ordered unless it is characteristic zero.
NOTE: yes, in George Leo Watson, Integral Quadratic Forms, pages 17-19, he discusses a specific version of this as being Hermite's method of reduction. Earlier on page 9 he mentions a more general version under the heading "Rational diagonalization" where he does not multiply through in order to always have integer entries.
Best with induction on $n.$ Re-order so that $a_{11} \neq 0.$
Consider
$$ a_{11}(x_1 + \frac{a_{12}}{a_{11}} x_2 + \frac{a_{13}}{a_{11}} x_3 + \cdots + \frac{a_{1n}}{a_{11}} x_n )^2 $$
This is
$$ a_{11} \left( x_1^2 + 2 \frac{a_{12}}{a_{11}} x_1 x_2 + 2 \frac{a_{13}}{a_{11}} x_1 x_3 + \cdots + 2 \frac{a_{1n}}{a_{11}} x_1 x_n + \mbox{ OTHER} \right)
$$
or
$$ a_{11} x_1^2 + 2 a_{12} x_1 x_2 + 2 a_{13} x_1 x_3 + \cdots + 2 a_{1n} x_1 x_n + \mbox{ OTHER}_2
$$
Now, subtract this off from the original $Q$ you wrote. What remains is a quadratic form in $n-1$ variables. Repeat the first step.
CAVEAT. I need to think about what happens when the only diagonal entries remaining are zero. We just make a trivial change to force the first diagonal entry to be nonzero. For example, consider the form $2xy$ in just two variables. If we take $x = u-v, \; y = u+v$ the new form is $2 u^2 - 2 v^2.$ That is, the ability to force the first diagonal coefficient to be nonzero is simply the ability for the form to represent a nonzero number, that is the form is not the zero form.