diff --git a/Measuring-performance.md b/Measuring-performance.md
new file mode 100644
index 0000000..0cbe250
--- /dev/null
+++ b/Measuring-performance.md
@@ -0,0 +1,54 @@
+It can be useful to know the performance impact of a code change in a real-world scenario (e.g. large codebase).
+
+The best way to do this is to record a fixed clangd session that you can replay, and extract timings from the event logs. Then you can run it against the old/new code, run it repeatedly to reduce noise etc.
+
+## Useful flags
+
+ - `--input-mirror-file=/tmp/mirror` will record editor->server messages, which can later be fed back into clangd. (Short sessions work best).
+ - `--input-style=delimited` makes it easier to edit these editor->server messages by hand
+ - injecting a `sync` message will force clangd to wait for everything (e.g. indexing) to finish before continuing
+ - running with env `CLANGD_TRACE=/tmp/trace.json` will emit machine-readable timings of various events annotated with `trace::Span`
+ - Flags like `-sync -background-index=0 -enable-config=0` improve determinism/isolation a bit
+
+## Example: measuring dynamic preamble indexing
+
+This is easy to trigger: we just need to open a file and wait for it to get indexed. The indexing happens during the "Running PreambleCallback" event which we want the duration of.
+
+First we create a basic clangd session script:
+```
+env CLANGD_FLAGS=-input-mirror-file=/tmp/mirror vim clang-tools-extra/clangd/XRefs.cpp
+```
+After opening and then closing vim, we look at the /tmp/mirror file, and convert it to delimited format (Replace the `Content-Length: NNN\n` at end of lines with `\n---\n`.
+
+If we replay it into clangd immediately we see from the logs that it exits too quickly, before actually indexing. So we add the `sync` message `{"jsonrpc":"2.0","id":999,"method":"sync","params":null}` at the end.
+
+Now we can replay this and produce a trace: 
+```
+env CLANGD_TRACE=/tmp/trace.json bin/clangd -sync --background-index=0 -pretty -input-style=delimited --enable-config=0 < /tmp/mirror
+```
+
+Looking at the trace file we can see the event we want:
+```
+  {
+    "pid": 0,
+    "ph": "X",
+    "name": "Running PreambleCallbacks",
+    "ts": 4283638.0920000002,
+    "tid": 485893,
+    "dur": 1216083.5929999994,
+    "args": {}
+  },
+```
+
+We can extract this with `jq`, also grabbing another event for reference:
+```
+jq -c '.traceEvents[] | select(IN(.name; "BuildPreamble", "Running PreambleCallback")) | {(.name): .dur}' < /tmp/trace.json
+```
+
+This produces the output we want:
+```
+{"Running PreambleCallback":1216083.5929999994}
+{"BuildPreamble":5523891.464}
+```
+
+Wrapping the last two steps in a script and running this several times on both the old & new code produces useful benchmark numbers.
\ No newline at end of file