diff --git a/doc/DesignConsiderations.md b/doc/DesignConsiderations.md
new file mode 100644
index 0000000..8e72eeb
--- /dev/null
+++ b/doc/DesignConsiderations.md
@@ -0,0 +1,121 @@
+Design Considerations
+=====================
+
+This document describes several of the factors that constrain the design of
+`libtock_core`.
+
+## Size impact
+
+Tock is designed to run on hardware with limited program storage and limited
+RAM. On Google's H1 chip, the Tock kernel and apps are limited to a 256 KiB
+flash region. The HiFive1 Rev B board has only 16 KiB of RAM.
+
+Note that a Tock system may have multiple process binaries, each of which uses
+`libtock_core`, so `libtock_core` must minimize its size impact to the extent
+possible.
+
+A process binary is generally stored in non-volatile memory and contains:
+
+  * Runtime headers (the `.crt0_header` section)
+  * Executable code (the `.text` section)
+  * Read-only data (the `.rodata` section)
+  * Non-zero-initialized read-write data (the `.data` section)
+
+A process' memory section is in RAM and contains:
+
+  * Heap memory (if a dynamic memory allocator is present)
+  * Zero-initialized read-write data (the `.bss` section)
+  * Non-zero-initialized read-write data (the `.data` section)
+  * Stack memory
+
+`libtock_core` developers should consider the space usage of their code, and
+where the space usage is (non-volatile memory or RAM). Important note: the
+`.data` section consumes space in both non-volatile memory and RAM!
+
+Most apps will not use all functionality in `libtock_core`. The size impact of
+using `libtock_core` in an app should be commensurate with the `libtock_core`
+functionality that app uses. If done correctly, this will allow `libtock_core`'s
+users to build Tock systems with multiple small, single-purpose apps without
+excessive code duplication.
+
+## Testability
+
+Programmatic tests are important to verify that new functionality works
+correctly and to keep existing functionality working through nontrivial
+refactorings. Different types of tests offer different capabilities:
+
+  * **Host-based unit tests** run on a non-Tock host OS (such as Linux). They
+    can only be used with portable `libtock_core` code, as the code they test
+    must run on both Tock and the host OS. These tests must either test code
+    that does not directly use system calls, or must direct the system calls to
+    a fake Tock kernel. These tests can be fast (compile and run in seconds),
+    and can easily simulate a variety of failure modes (e.g. kernel errors) that
+    may be difficult to generate with a real Tock kernel.
+  * **Emulation tests** run on an emulated Tock system. Currently, emulation
+    tests run in QEMU on an emulated HiFive1 Rev B board. These test
+    `libtock_core`'s interaction with a real Tock kernel, albeit with limited
+    hardware access.
+  * **Hardware tests** run on physical hardware with a real Tock kernel. These
+    are end-to-end tests that test not only `libtock_core`'s functionality but
+    also the Tock kernel's interaction with the real hardware. However,
+    automating these tests is difficult, and currently we do not have a way to
+    run hardware tests in CI.
+
+If `libtock_core` is functional on real hardware with a real kernel, it should
+be able to support emulation tests and hardware tests. However, host-based unit
+tests require more work. In its intended use case, `libtock_core` is not
+supposed to be portable, supporting only the Tock kernel. Hoever, host-based
+unit tests are very valuable, both for rapid development and for testing error
+handling code. To support host-based tests, `libtock_core` must be built out of
+portable pieces that can be unit tested, even though the library as a whole is
+not portable.
+
+Testing pieces of `libtock_core` against a fake kernel can be done using
+dependency injection, but most dependency injection techniques have considerable
+code size and RAM usage costs. `libtock_core` needs either a dependency
+injection technique that has minimal size impact or an alternative mechanism for
+testing `libtock_core`'s system call error handling logic.
+
+## Support multiple asynchronous programming models.
+
+The general Rust ecosystem has converged on futures as its building block for
+interoperable asynchronous APIs, but futures have a [size
+cost](https://github.com/tock/design-explorations/tree/master/size_comparison)
+that makes them impractical for some use cases of `libtock_core`. Although
+`libtock_core` is still a work in progress (see [issue
+217](https://github.com/tock/libtock-rs/issues/217)), it will probably use an
+asynchronous API design based on the [Asynchronous Components using Zero Sized
+Type
+Pointers](https://github.com/tock/design-explorations/tree/master/zst_pointer_async)
+exploration.
+
+In order to interoperate with other libraries as well as the `async`/`await`
+functionality built in to Rust, we need `libtock_core` to interface well with
+futures. In addition, there are several use cases of `libtock_core` that consist
+primarily of synchronous code, so we need to be able to use `libtock_core`'s
+APIs in a synchronous manner.
+
+The `libtock-c` developers are already familiar with some of the complications
+that come up when synchronous code is interfaced to asynchronous code. For
+example, calling `printf` -- an extremely common debugging tool -- causes
+callbacks to run in the background. This causes surprising and hard-to-debug
+reentrancy issues.
+
+## Dependency tree minimization
+
+Some of users of `libtock_core` intend to use it in applications that will
+undergo code audits for security certifications. To make this auditing
+practical, it is important to allow users to use `libtock_core` without pulling
+in a large dependency tree.
+
+As a general rule, a `libtock_core` should avoid having required dependencies
+unless those dependencies are minimal (that is, avoiding the dependency would
+require reimplementing nearly all of it). Dependencies included in the Rust
+toolchain (such as the `core` crate and its dependencies) are an exception, as
+they are a part of the language itself.
+
+## No hard `alloc` dependency
+
+It should be possible to use `libtock_core` without bringing in the `alloc`. Of
+course, it is fine for `libtock_core` to offer a memory allocator as an optional
+feature.