diff --git a/.github/copilot-instructions.md b/.github/copilot-instructions.md
new file mode 100644
index 0000000..c4f44ee
--- /dev/null
+++ b/.github/copilot-instructions.md
@@ -0,0 +1,72 @@
+# Copilot Instructions — ramanujantools
+
+You are an AI developer working on `ramanujantools`, a Python library for
+symbolic and numeric matrix operations used by the
+[Ramanujan Machine](https://www.ramanujanmachine.com/) project.
+
+## Project overview
+
+`ramanujantools` provides tools for Conservative Matrix Fields (CMFs),
+Polynomial Continued Fractions (PCFs), linear recurrences, and high-precision
+constant computation. The numeric backend uses
+[python-flint](https://github.com/flintlib/python-flint) for exact rational
+arithmetic.
+
+## Architecture
+
+- `ramanujantools/matrix.py` — `Matrix` (SymPy-based), public API entry point.
+- `ramanujantools/flint_core/numeric_matrix.py` — `NumericMatrix` (FLINT `fmpq_mat` wrapper).
+- `ramanujantools/flint_core/symbolic_matrix.py` — `SymbolicMatrix`.
+- `Matrix.walk()` dispatches to `NumericMatrix.walk()` or `SymbolicMatrix.walk()` based on input type.
+- `@batched` decorator (`ramanujantools/utils/batched.py`) allows `iterations` to be scalar or list.
+- `NumericMatrix.lambda_from_rt()` compiles a SymPy matrix into a fast FLINT evaluator.
+- `Matrix._walk_inner` has an LRU cache; clear with `Matrix._walk_inner.cache_clear()` before benchmarks.
+
+## Code conventions
+
+### Testing
+- Tests live next to source as `*_test.py` files (e.g., `numeric_matrix_test.py`).
+- Run with `pytest`. Benchmarks use `pytest-benchmark`.
+- Benchmarks live in `matrix_benchmark.py` and run **strictly via pytest**, never as standalone scripts.
+- Every new public function needs at least one test (happy path + edge case).
+- Tests should exercise the public `Matrix` API, not internal classes directly, unless testing internal-only behavior.
+- Use `@pytest.mark.parametrize` to cover multiple inputs (e.g., different matrix sizes, depths, edge cases) in a single test function instead of duplicating test functions.
+
+### Numeric precision
+- **Never use Python `float`** for mathematical verification. Use `mpmath.mpf`, `sympy.Rational`, or FLINT types.
+- Set `mpmath.mp.dps` to at least 2× the digits you need.
+- Verify formulas to 100+ decimal places.
+
+### Style
+- Clear, descriptive variable names (e.g., `first`/`last` not `lo`/`hi`).
+- Extract magic numbers into named constants (e.g., `SEQUENTIAL_THRESHOLD = 8`).
+- Docstrings on public functions: one-line summary, parameters, return value.
+- No unnecessary abstractions for one-off operations.
+- **Minimal comments.** Do not add decorative section separators (`# -----`), redundant header comments, or obvious comments. Let the code speak for itself. A brief comment is fine when explaining *why*, not *what*.
+
+### Performance
+- Profile before optimizing (`cProfile`, `timeit`).
+- Benchmark before and after, at realistic scale (N ≥ 1000 for walks).
+- Binary splitting (`_product_tree`) is used in `NumericMatrix.walk()` — balanced divide-and-conquer is critical for FLINT rationals where entry sizes grow with each multiplication.
+
+## PR review workflow
+
+When responding to PR review comments:
+1. Read every review comment carefully before making changes.
+2. Prefer the simplest fix that addresses the reviewer's concern.
+3. Run the relevant tests after every change (`pytest <test_file> -v`).
+4. If the reviewer asks to check for existing utilities/packages, do the research and report findings even if no suitable alternative exists.
+5. Commit with a clear message referencing which review comments are addressed.
+
+## Reviewer preferences (RotemKalisch)
+
+- Benchmarks must be pytest-only — no `if __name__ == "__main__"` standalone scripts.
+- Test through the public `Matrix` class API where possible.
+- Check for existing pythonic utilities/packages before writing custom algorithms.
+- Justify any new dependency with a comparison of how the code looks and performs with vs. without it.
+
+## Repository rules
+
+- Do not push to `master` directly. Use feature branches and PRs.
+- Run the full test suite (`pytest`) before pushing.
+- The CI runs `python-package.yml` — ensure it passes.