diff --git a/src/assemblers/README.md b/src/assemblers/README.md
new file mode 100644
index 0000000..41334d9
--- /dev/null
+++ b/src/assemblers/README.md
@@ -0,0 +1,27 @@
+# src/assemblers/
+
+Assembly machinery for the current type-stable pipeline (0.x).
+
+See `AGENTS.md` for invariants and `docs/src/assembler_choice.md` for a short
+comparison of DOF-based, element-based, and matrix-free paths. Regression tests
+live under `test/assemblers/`, especially `test_dof_based_zero_alloc.jl` and
+`test_matrix_free_operator.jl`.
+
+The directory is split along the real fault lines. Each subdirectory
+owns a specific assembly strategy plus the supporting types it
+specialises:
+
+| Path                     | Responsibility                                                                                                       |
+|--------------------------|----------------------------------------------------------------------------------------------------------------------|
+| `abstract.jl`            | `AbstractAssembler` hierarchy plus the kernel defaults (`get_field`, `dofs_per_node`, `get_dof_mapping!`).           |
+| `microkernel.jl`         | Microkernel trait (`evaluate_entry`, `qpoint_buffer_eltype`, ...) used by the DOF-based and matrix-free path.        |
+| `caches/`                | Reusable per-element scratch (`GeometryCache`, `ElementCache`, `AssemblyMaterialWorkspace`) and the COO global cache.|
+| `element_based/`         | Classic element-by-element assembler (COO) plus the scatter routines it dispatches to.                               |
+| `dof_based/`             | DOF-by-DOF assembler (`DOFBasedCOOAssembler` / `DOFBasedCOOCache` and the backend-agnostic KernelAbstractions port). |
+| `matrix_free/`           | Declarative Dirichlet / MPC constraints, Neumann loads, matrix-free preconditioners and the generalized eigensolver. |
+
+The matrix-free directory builds on the DOF-based cache plus
+`apply_K!` / `apply_M!`. All matrix-free constraint and load types
+share the `apply_constraint_*` hook protocol so that Dirichlet, MPC
+and load contributions compose freely with both the assembled and the
+matrix-free solves.