diff --git a/doc/contributing.rst b/doc/contributing.rst
index 045ec2cc..b4b94bba 100644
--- a/doc/contributing.rst
+++ b/doc/contributing.rst
@@ -13,6 +13,11 @@ Contributing
 
 This section will be expanded later. For now it lists docstrings for a number of internal variables, classes and functions.
 
+.. toctree::
+   :maxdepth: 1
+
+   releasing
+
 Variables
 ~~~~~~~~~
 
diff --git a/doc/releasing.rst b/doc/releasing.rst
new file mode 100644
index 00000000..5da39fbc
--- /dev/null
+++ b/doc/releasing.rst
@@ -0,0 +1,85 @@
+.. _releasing:
+
+Releasing cf-xarray
+===================
+
+This checklist is for maintainers with upstream release permissions.
+The existing workflows, rather than a local upload, build and publish releases.
+The version comes from Git tags through ``setuptools_scm``; do not manually
+edit a source version string.
+
+Before tagging
+--------------
+
+* Confirm the intended changes are merged into ``main`` and the required CI
+  jobs pass. Review the changes since the previous release and prepare release
+  notes using the `GitHub releases page
+  <https://github.com/xarray-contrib/cf-xarray/releases>`_.
+* Check the documentation build and any compatibility changes against the
+  supported Python and dependency environments in
+  ``.github/workflows/ci.yaml``.
+* For a package-build rehearsal, a maintainer can add the ``test-build`` label
+  to a pull request or manually dispatch ``testpypi-release.yaml``. Despite its
+  name, that workflow currently builds and checks the distributions and installs
+  the wheel, but its TestPyPI upload step is commented out. Inspect its
+  ``releases`` artifact and the version printed by the installed package.
+  An ordinary push to ``main`` does not pass this job's execution condition.
+
+Local build rehearsal
+---------------------
+
+Use a full Git checkout with tags available. Run these commands in a disposable
+checkout and a fresh virtual environment, not in an environment with an editable
+cf-xarray installation:
+
+.. code-block:: bash
+
+   python -m pip install build twine
+   python -m build
+   python -m twine check --strict dist/*
+   python -m pip install dist/cf_xarray-*.whl
+
+Then change to a directory outside the checkout and check the installed wheel:
+
+.. code-block:: bash
+
+   python -m cf_xarray.scripts.print_versions
+   python -c "import cf_xarray; print(cf_xarray.__version__)"
+
+Use a clean ``dist`` directory so the install selects exactly one wheel.
+A development checkout can have a development version; the final tagged build
+must have the intended release version, not ``0.0.0``.
+This rehearsal does not upload anything. The CI workflows clean their disposable
+checkouts before building; do not copy ``git clean -xdf`` into a working checkout
+that contains files you want to keep.
+
+Tag and publish
+---------------
+
+* After the checks pass, choose the release version and create and push its
+  ``v``-prefixed Git tag at the reviewed commit using the project's normal
+  maintainer process.
+* A pushed ``v*`` tag triggers ``pypi.yaml`` in the upstream repository. Its
+  build job creates and checks the sdist and wheel; the push-event job installs
+  the wheel, prints versions and uploads to TestPyPI. Confirm these jobs succeed
+  before publishing the GitHub Release. A TestPyPI upload is a real publication,
+  not a dry run.
+* Publish the GitHub Release for the same tag, with the reviewed release notes.
+  The ``release: published`` event triggers ``pypi.yaml`` again and publishes to
+  PyPI using the protected ``pypi`` environment and trusted publishing.
+  Complete any required environment approval and confirm the upload succeeded.
+  Do not recreate or move a published tag, or attempt to reuse a published
+  package version to recover from a failed upload.
+
+After publishing
+----------------
+
+* In a fresh environment, install the released version from PyPI and run
+  ``python -m cf_xarray.scripts.print_versions`` outside the source checkout.
+  Check the installed version matches the GitHub Release and both distribution
+  files are present on PyPI.
+* Check that the documentation for the release is available and that the release
+  notes describe compatibility changes and credit contributors.
+* If a release is broken, preserve the failed workflow logs and artifacts,
+  coordinate the correction with the maintainers, and use a new version rather
+  than silently replacing the release.