A spectral sequence is a systematic array of pages E_{p,q}^r with differentials d^r: E_{p,q}^r → E_{p-r,q+r-1}^r, organized by a filtration. Successive pages E^{r+1} are the homology of differentials on E^r, and the sequence converges to the associated graded of a target complex. Spectral sequences arise from filtered chain complexes, double complexes, and fibrations, providing powerful computational tools for homology that break hard problems into successively finer approximations.
Study the long exact sequence as a degenerate spectral sequence. Compute homology of the total complex of a double complex via spectral sequences. Apply the Serre spectral sequence to compute homology of fibration total spaces from base and fiber homology.
Spectral sequences compute the associated graded of the target, not the target directly; loss of information via filtration requires care. Differentials on higher pages depend on previous pages non-trivially; simply knowing early pages does not determine the full sequence. Convergence is a separate condition and can fail if the filtration is non-bounded or degenerates.
From your prerequisites, you know that a chain complex (C_*, d) has homology groups H_n(C) measuring cycles that are not boundaries, and that exact sequences encode algebraic relationships between homology groups of different spaces. Spectral sequences generalize both: they are a systematic machine for computing homology of a complex when the complex has additional structure — a filtration — that lets you attack the computation in layers rather than all at once.
A filtration of a chain complex C is a nested sequence of subcomplexes: ... ⊆ F_{p-1}C ⊆ F_pC ⊆ F_{p+1}C ⊆ ... ⊆ C. The filtration breaks C into layers; each quotient F_pC/F_{p-1}C is a "slice" of the complex. The idea is that homology of these slices is easier to compute than homology of C directly, and a spectral sequence systematically tracks how those slice-level computations assemble into the full answer. The E² page (or E¹ in some conventions) is the array of homology groups of the associated graded complex — one entry E_{p,q} for each bidegree (p,q) where p tracks the filtration level and q tracks the complementary degree.
The mechanism is iterated approximation. The E^r page (the r-th page) consists of groups E_{p,q}^r together with differentials d^r: E_{p,q}^r → E_{p-r, q+r-1}^r that shift filtration degree by −r and total degree by +1. The E^{r+1} page is the homology of d^r: E_{p,q}^{r+1} = ker(d^r)/im(d^r). As r increases, successive differentials kill off groups that are "boundaries at the r-th order of approximation" and reveal the next layer of structure. For most applications, the differentials eventually vanish (d^r = 0 for all large r), and the sequence converges: E_{p,q}^∞ is the associated graded of the filtration on H_{p+q}(C). Think of it as repeatedly zooming in — each page refines your knowledge of the homology until nothing more is hidden.
A powerful concrete example: the Serre spectral sequence for a fibration F → E → B, where E is the total space, B the base, and F the fiber. Here the E² page has E_{p,q}² = H_p(B; H_q(F)), the homology of the base with coefficients in the homology of the fiber. The spectral sequence converges to H_*(E). If you know the homology of B and F, you can often compute H_*(E) by tracking which differentials d^r are nonzero. For example, the Hopf fibration S¹ → S³ → S² immediately tells you (via the Serre spectral sequence) what the differentials must be, recovering the homology of S³ from knowledge of S¹ and S² — a calculation that would be harder by direct methods.
The critical subtlety your misconceptions section raises is that convergence gives you the associated graded of H_*(C), not H_*(C) itself. There can be extension problems: even knowing all the associated graded pieces E_{p,q}^∞, there may be multiple non-isomorphic groups H_n(C) with that associated graded. (This is the same issue as knowing a group's composition factors without knowing the extension.) Spectral sequences with integer coefficients can have extension problems; spectral sequences with field coefficients do not, because every short exact sequence of vector spaces splits. In practice, working over a field eliminates extensions and makes spectral sequence computations fully algorithmic — which is why topology courses often introduce them over ℤ/2 or ℚ before tackling integer coefficients.