diff --git a/docs/REFERENCE.md b/docs/REFERENCE.md
new file mode 100644
index 0000000..a44046d
--- /dev/null
+++ b/docs/REFERENCE.md
@@ -0,0 +1,104 @@
+# REFERENCE.md — GitHub Issue RFC Guide
+
+Used by Claude and developers when writing architectural refactor issues for this repository.
+
+## Skills that use this file
+
+- [`improve-codebase-architecture`](https://github.com/mattpocock/skills/tree/main/improve-codebase-architecture) — surface architectural friction and propose deep-module refactors as GitHub issue RFCs
+- `gsd` suite — built-in Claude Code workflow tool for structured project planning, phase execution, and delivery
+
+---
+
+## GitHub Issue RFC Template
+
+Use this structure for all refactor / architecture RFC issues:
+
+```markdown
+## Why
+[What creates the friction and why it matters technically.
+Cover: which modules are affected, root cause of the friction
+(coupling, shallow interface, silent failure, poor testability),
+and the concrete technical consequence if left unaddressed.]
+
+## Current behaviour
+[A real code snippet — not prose — showing the problem.]
+
+## Expected outcome
+[Bulleted vision of the module after the refactor.
+Not a task list — describe the end state.]
+
+## Acceptance criteria
+- [ ] Specific, verifiable requirement
+- [ ] Each item can be checked by a reviewer without knowing the implementation
+- [ ] Include test requirements explicitly
+```
+
+### Section guidance
+
+**Why** must answer:
+- Which file(s) and line(s) are the friction source?
+- What design principle is violated (coupling, shallow module, implicit contract, etc.)?
+- What breaks or degrades as a result (testability, observability, maintainability)?
+
+**Current behaviour** must show code, not describe it. One focused snippet is better than many.
+
+**Expected outcome** describes the *shape* of the solution, not the steps. A reader should be able to evaluate it without reading the implementation plan.
+
+**Acceptance criteria** are checkbox items. Each must be independently verifiable. Include at least one test-related criterion.
+
+---
+
+## Four Dependency Categories
+
+Use these when classifying the coupling between modules in a candidate refactor:
+
+| Category | Description | Example |
+|----------|-------------|---------|
+| **In-process** | Pure computation, in-memory state, no I/O. Always deepenable — merge modules and test directly. | `template_registry.py` ↔ `template_model.py` within `templates/` |
+| **Local-substitutable** | Dependencies with local test stand-ins. Deepenable if the substitute exists. | `TemplateRegistry` reading JSON from disk — testable with `tmp_path` |
+| **Remote but owned (Ports & Adapters)** | Your own services across a network boundary. Define a port (interface); inject an in-memory adapter for tests, real adapter in production. | Internal APIs or services you control |
+| **True external (Mock)** | Third-party services you don't control. Mock at the boundary; inject as a port. | External auth providers, third-party APIs |
+
+---
+
+## Dependency Strategy Patterns
+
+How to handle each category when designing a deep module:
+
+| Category | Strategy |
+|----------|----------|
+| **In-process** | Merge modules; test the combined boundary directly |
+| **Local-substitutable** | Test with the local stand-in running in-process (e.g. `tmp_path`, in-memory store) |
+| **Remote but owned** | Define a Protocol port; production uses real adapter, tests use in-memory adapter |
+| **True external** | Define a Protocol port; inject a mock in tests |
+
+---
+
+## Deep Module Checklist
+
+A module is **deep** (Ousterhout, *A Philosophy of Software Design*) when its interface is significantly simpler than its implementation. Before raising a refactor RFC, verify the friction is real:
+
+- [ ] Must a caller read the *implementation* (not just the signature) to use the module correctly?
+- [ ] Does the module leak I/O, error handling, retry logic, or data transformation into its interface?
+- [ ] Are dependencies hardcoded or constructed internally, making the module untestable without patching?
+- [ ] Does the module raise built-in exception types from *its dependencies* rather than its own typed exceptions?
+- [ ] Is there more than one construction path, or a multi-branch constructor choosing between behaviours?
+
+If any of these is yes, the module is a candidate.
+
+A module is **shallow** (and worth consolidating, not splitting) when:
+- Its interface has nearly as many concepts as its implementation
+- Tests mock internals rather than testing at the boundary
+- An abstraction (Protocol, base class) exists with only one implementation and no planned others
+
+---
+
+## Worked Examples
+
+Issues #62–#64 on this repository are RFCs written using this format. They cover:
+
+- [#62](https://github.com/longieirl/bankstatementprocessor/issues/62) — hardcoded Euro symbol in output column names
+- [#63](https://github.com/longieirl/bankstatementprocessor/issues/63) — currency-specific internal TypedDict field names
+- [#64](https://github.com/longieirl/bankstatementprocessor/issues/64) — inconsistent currency-symbol stripping across code paths
+
+Use them as style reference.