diff --git a/README.md b/README.md
index 70341f6..80def30 100644
--- a/README.md
+++ b/README.md
@@ -39,3 +39,51 @@ arguments cannot be passed to a Makefile target.
   -workers int
         number of workers to use (default 2*<number-of-cores>)
 ```
+
+# Approach
+
+To satisfy the project requirements, the following approach was used:
+
+1. Order files by timestamp, to be able to step through the "evolution" of the
+   associated documents.
+2. Iterate through all the timestamps, and for all documents identified,
+   determine whether any of the files associated with the current timestamp can be
+   related to an existing document. For each file that remains unassociated with an
+   existing document at this point, create a new document.
+3. Sort the associated files for each document in order by their filename (which
+   is a number).
+
+## Practical considerations
+
+The corpus provided for this assignment is comprised of 20,000 files; a
+substantial enough amount to take time and memory complexity into account. To
+manage time complexity, a pool of workers compares documents against candidate
+files in a parallel manner, rather than serially stepping through all the files.
+On my MacBook Pro M1 Max, this reduces runtime from 27 minutes to roughly five
+minutes. Additionally, file contents are cached to prevent repeatedly reading
+the same files off disk.
+
+Cached files are purged as soon as their contents won't ever be needed again, to
+conserve a large amount of memory that would be used by caching the contents of
+20,000 files. We can be sure a file's contents aren't needed anymore when its
+associated timestamp has already been processed, and the file isn't the most
+recent file associated with a given document: we only need the latest version of
+the document to associate files from future timestamps.
+
+There is some status text written to the console by default, but it is written
+to stderr, so if only the pure list of files associated with one another is
+desired, redirecting stdout will yield a clean list without any status text. The
+same can be accomplished by using the `-output` flag, as well.
+
+# High level design
+
+The main representational types in the program are the `DocumentManager` and the
+`Document` types. `DocumentManager` handles starting workers and comparing
+documents against candidate files. The `Document` type keeps track of which
+files are associated with it, and the timestamp of the latest file. Work is
+performed by sending "work items" to the workers on a channel, who pull a work
+item off and perform it, altering documents if they need to be updated with a
+new associated file. When all files have been checked, the final set of
+documents is presented to the user.
+
+![High level design](design.png)
diff --git a/design.png b/design.png
new file mode 100644
index 0000000..ce8a983
Binary files /dev/null and b/design.png differ