diff --git a/docs/design/bit.md b/docs/design/bit.md
new file mode 100644
index 0000000..9340ce6
--- /dev/null
+++ b/docs/design/bit.md
@@ -0,0 +1,63 @@
+# bit
+
+Source code: https://github.com/intel/cpp-std-extensions/blob/main/include/stdx/bit.hpp  
+Documentation: https://intel.github.io/cpp-std-extensions/#_bit_hpp
+
+## No more manual bit twiddling!
+
+A lot of the functionality in this header exists to put an end to manual bit
+shifting, masking, etc. Yes it's easy, but it's not simple. Manual bit
+operations are prone to sign and integer promotion errors, and sprinkling
+`static_cast` everywhere to guard against this makes things ugly and less
+understandable.
+
+## `to_be`, `from_be`, `to_le`, `from_le`
+
+`to_be` and `from_be` have identical implementations. As do `to_le` and
+`from_le`. But they are different in showing the intent of user code. And also
+different in pre and post conditions:
+
+```cpp
+// pre: x is big-endian
+// post: x is platform-endian
+auto x = stdx::from_be(y);
+
+// pre: x is platform-endian
+// post: x is big-endian
+auto x = stdx::to_be(y);
+```
+
+In other words, the use of `from_be`/`to_be` is very different from the use of
+`std::byteswap` in cross-platform friendly code.
+
+## `bit_pack` and `bit_unpack`
+
+These functions have some interesting endianness concerns. It is important that
+the functions round-trip properly.
+
+```cpp
+std::uint8_t a{0x12u}, b{0x34u}, c{0x56u}, d{0x78u};
+auto x = stdx::bit_pack<std::uint32_t>(a, b, c, d);
+auto [a_, b_, c_, d_] = stdx::bit_unpack<std::uint8_t>(x);
+// a == a_, b == b_; c == c_, d == d_
+```
+
+This also means that that "written order" (big-endian) is the expected ordering.
+
+```cpp
+auto [a, b, c, d] = stdx::bit_unpack<std::uint8_t>(0x12345678u);
+// a == 0x12, b == 0x34; c == 0x56, d == 0x78
+```
+
+## `bit_destructure`
+
+`bit_destructure` and `bit_unpack` are similar, but where `bit_unpack` is
+"big-endian" (for the above reasons), `bit_destructure` is "little-endian".
+Because it makes sense to count **up** in bits.
+
+Also, N split points mean N+1 values.
+
+```cpp
+auto [a, b, c, d] = stdx::bit_destructure<8, 16, 24>(0x12345678u);
+// a == 0x78, b == 0x56; c == 0x34, d == 0x12
+```
diff --git a/docs/design/cached_latched.md b/docs/design/cached_latched.md
new file mode 100644
index 0000000..df378de
--- /dev/null
+++ b/docs/design/cached_latched.md
@@ -0,0 +1,29 @@
+# cached and latched
+
+Source code: https://github.com/intel/cpp-std-extensions/blob/main/include/stdx/latched.hpp  
+Documentation: https://intel.github.io/cpp-std-extensions/#_latched_hpp
+
+Source code: https://github.com/intel/cpp-std-extensions/blob/main/include/stdx/cached.hpp  
+Documentation: https://intel.github.io/cpp-std-extensions/#_latched_hpp
+
+## Lazy values
+
+Both `cached` and `latched` model lazy computation of values. Under the hood
+they are just `std::optional`. 
+
+Note that use of `cached` and `latched` generally incurs a branch check for each
+use.
+
+`latched` does not provide for `reset`/`refresh`: it's intended for values that
+are read once at startup time.
+
+`cached` values can be `reset` or `refresh`ed. The default should probably be to
+`reset` and lazily recompute but `refresh` is for the use case of "recompute
+now".
+
+### Other ideas
+
+We don't use `stdx::optional` here; that requires tombstone value(s). Maybe
+there is a way to select the optional implementation based on detecting a
+tombstone? in practice many use cases for `cached`/`latched` are register
+(`std::uint32_t`) reads rather than strong types that could use tombstones.
diff --git a/docs/design/utility.md b/docs/design/utility.md
new file mode 100644
index 0000000..b65a063
--- /dev/null
+++ b/docs/design/utility.md
@@ -0,0 +1,24 @@
+# utility
+
+Source code: https://github.com/intel/cpp-std-extensions/blob/main/include/stdx/utility.hpp  
+Documentation: https://intel.github.io/cpp-std-extensions/#_utility_hpp
+
+## sized
+
+This exists so that we don't have to reason about "raw" size calculations
+everywhere:
+
+```cpp
+// not this
+auto size_in_dwords = (size_in_bytes + 3) / 4;
+// but this
+auto size_in_dwords = stdx::sized8{size_in_bytes}.in<std::uint32_t>();
+```
+
+More verbose, but almost certainly less error-prone and easier to read. These
+are simple operations, and sometimes they should be terse. If the values are
+known at compile time:
+
+```cpp
+constexpr auto size_in_dwords = 42_z8->z32;
+```