diff --git a/src/documentation/safety-comments.md b/src/documentation/safety-comments.md
new file mode 100644
index 0000000..5cbe19d
--- /dev/null
+++ b/src/documentation/safety-comments.md
@@ -0,0 +1,79 @@
+# Safety comments
+
+Using [`unsafe`] blocks is often required in the Rust compiler or standard
+library, but this is not done without rules: each `unsafe` block should have
+a `SAFETY:` comment explaining why the block is safe, which invariants are
+used and must be respected. Below are some examples taken from the standard
+library:
+
+[`unsafe`]: https://doc.rust-lang.org/stable/std/keyword.unsafe.html
+
+## Inside `unsafe` elements
+
+This one shows how an `unsafe` function can pass the requirements through to its
+caller with the use of documentation in a `# Safety` section while still having
+more invariants needed that are not required from callers. `clippy` has a
+lint for `# Safety` sections by the way.
+
+[See the example on github][as_bytes_mut]
+
+```rust
+/// Converts a mutable string slice to a mutable byte slice.
+///
+/// # Safety
+///
+/// The caller must ensure that the content of the slice is valid UTF-8
+/// before the borrow ends and the underlying `str` is used.
+///
+/// Use of a `str` whose contents are not valid UTF-8 is undefined behavior.
+///
+/// ...
+pub unsafe fn as_bytes_mut(&mut self) -> &mut [u8] {
+    // SAFETY: the cast from `&str` to `&[u8]` is safe since `str`
+    // has the same layout as `&[u8]` (only libstd can make this guarantee).
+    // The pointer dereference is safe since it comes from a mutable reference which
+    // is guaranteed to be valid for writes.
+    unsafe { &mut *(self as *mut str as *mut [u8]) }
+}
+```
+
+This example is for a function but the same principle applies to `unsafe trait`s
+like [`Send`] or [`Sync`] for example, though they have no `# Safety` section
+since their entire documentation is about why they are `unsafe`.
+
+Note that in the Rust standard library, [`unsafe_op_in_unsafe_fn`] is active
+and so each `unsafe` operation in an `unsafe` function must be enclosed in an
+`unsafe` block. This makes it easier to review such functions and to document
+their `unsafe` parts.
+
+[`Send`]: https://doc.rust-lang.org/stable/std/marker/trait.Send.html
+[`Sync`]: https://doc.rust-lang.org/stable/std/marker/trait.Sync.html
+[as_bytes_mut]: https://github.com/rust-lang/rust/blob/a08f25a7ef2800af5525762e981c24d96c14febe/library/core/src/str/mod.rs#L278
+[`unsafe_op_in_unsafe_fn`]: https://doc.rust-lang.org/rustc/lints/listing/allowed-by-default.html#unsafe-op-in-unsafe-fn
+
+## Inside *safe* elements
+
+Inside safe elements, a `SAFETY:` comment must not depend on anything from the
+caller beside properly constructed types and values (i.e, if your function
+receives a reference that is unaligned or null, it is the caller fault if it
+fails and not yours).
+
+`SAFETY` comments in *safe* elements often rely on checks that are done before
+the `unsafe` block or on type invariants, like a division by `NonZeroU8` would
+not check for `0` before dividing.
+
+[See the example on github][split_at]
+
+```rust
+pub fn split_at(&self, mid: usize) -> (&str, &str) {
+    // is_char_boundary checks that the index is in [0, .len()]
+    if self.is_char_boundary(mid) {
+        // SAFETY: just checked that `mid` is on a char boundary.
+        unsafe { (self.get_unchecked(0..mid), self.get_unchecked(mid..self.len())) }
+    } else {
+        slice_error_fail(self, 0, mid)
+    }
+}
+```
+
+[split_at]: https://github.com/rust-lang/rust/blob/a08f25a7ef2800af5525762e981c24d96c14febe/library/core/src/str/mod.rs#L570
diff --git a/src/documentation/summary.md b/src/documentation/summary.md
index 1fb6e93..686bcac 100644
--- a/src/documentation/summary.md
+++ b/src/documentation/summary.md
@@ -1 +1,2 @@
 - [doc alias policy](./doc-alias-policy.md)
+- [safety comments policy](./safety-comments.md)