diff --git a/backends/apple/coreai/.llms/skills/code-writing.md b/backends/apple/coreai/.llms/skills/code-writing.md
new file mode 100644
index 00000000000..fabf66b7bec
--- /dev/null
+++ b/backends/apple/coreai/.llms/skills/code-writing.md
@@ -0,0 +1,54 @@
+---
+name: coreai-code-writing
+description: >-
+  Conventions for writing code in the Core AI ExecuTorch backend
+  (backends/apple/coreai): comment style (clear, concise, self-documenting) and
+  how to run the test suite. Apply when editing or adding code/tests here.
+---
+
+# Core AI backend: code-writing conventions
+
+## Comments and docstrings
+
+Write self-documenting code first; reach for a comment only when the code cannot
+speak for itself.
+
+- **Explain WHY, not WHAT.** Well-named identifiers already say what the code
+  does. Comment only non-obvious reasons: a hidden constraint, a subtle
+  invariant, a workaround, or behavior that would surprise a reader.
+- **Be clear and concise.** One short line is usually enough. Prefer a plain
+  sentence over a paragraph. Delete a comment if removing it wouldn't confuse a
+  future reader.
+- **No LLMisms.** Do not use em-dashes (`—`) or ` -- ` as sentence separators;
+  use periods, commas, semicolons, or parentheses. Do not use decorative dash
+  banners (`# --- section ---`); a plain `# Section.` line is enough. Avoid
+  markdown emphasis (`*x*`, `**x**`) and arrows (`=>`) in prose.
+- **Don't restate the task or the diff.** No "added for X", "used by Y", or
+  "handles the case from T123"; that belongs in the diff description and rots.
+- **Keep comments true.** When you change code, update or delete any comment it
+  affects rather than leaving a stale claim.
+
+Legitimate non-prose dashes are fine: CLI flags (`--platform`,
+`--min-deployment-version`) and shell/code tokens.
+
+## Running the tests
+
+Run the whole backend suite (both `test/` and `passes/test/`) with the helper
+script, from inside the Core AI conda env:
+
+```bash
+conda run -n coreai backends/apple/coreai/run_all_tests.sh
+```
+
+Extra args are forwarded to `unittest`, e.g. `... run_all_tests.sh -v` or
+`... run_all_tests.sh -k sidecar`.
+
+**Use this script (unittest), not `pytest` directly.** pytest's default
+discovery puts `backends/apple/` on `sys.path`, so `import coreai` resolves to
+this backend directory and shadows the real Apple `coreai` SDK (`coreai_torch`
+then fails importing `coreai.authoring`). unittest imports via the full
+`executorch.backends.apple.coreai.*` path and avoids the shadowing.
+
+The real-toolchain AOT test (`CoreAIAOTCompileTest`) is gated on
+`xcrun coreai-build`; it runs on macOS with the Metal Toolchain and skips
+elsewhere. The mocked AOT tests always run.
diff --git a/backends/apple/coreai/run_all_tests.sh b/backends/apple/coreai/run_all_tests.sh
new file mode 100755
index 00000000000..c7dd1b59a62
--- /dev/null
+++ b/backends/apple/coreai/run_all_tests.sh
@@ -0,0 +1,27 @@
+#!/usr/bin/env bash
+# Dev convenience: run all Core AI backend tests (test/ + passes/test/).
+#
+# WIP/under-construction backend -- a local helper, not a landing artifact.
+#
+# Uses unittest (NOT pytest): pytest's default discovery puts backends/apple/ on
+# sys.path, so `import coreai` resolves to THIS backend dir and shadows the real
+# Apple `coreai` SDK (coreai_torch then fails to import coreai.authoring).
+# unittest imports via the full executorch.backends.apple.coreai.* path, so the
+# SDK is not shadowed.
+#
+# Run inside the Core AI conda env, e.g.:
+#     conda run -n coreai backends/apple/coreai/run_all_tests.sh
+# Extra args are forwarded to unittest, e.g. `... run_all_tests.sh -v`.
+set -euo pipefail
+
+cd "$(sl root 2>/dev/null || git rev-parse --show-toplevel)"
+
+# Discover every test_*.py under the backend and turn it into a dotted module
+# name (executorch.backends.apple.coreai...<file without .py>).
+modules=$(
+  find backends/apple/coreai -name 'test_*.py' \
+    | sed -E 's#^#executorch.#; s#/#.#g; s#\.py$##' \
+    | sort
+)
+
+exec python -m unittest ${modules} "$@"