diff --git a/pythonetc/README.md b/pythonetc/README.md
index c13a2a9..977ad04 100644
--- a/pythonetc/README.md
+++ b/pythonetc/README.md
@@ -70,6 +70,7 @@ More:
 1. ./json-default.md
 1. ./ipython.md
 1. ./array.md
+1. ./re_compile.md
 
 Out of order:
 
diff --git a/pythonetc/re_compile.md b/pythonetc/re_compile.md
new file mode 100644
index 0000000..77e7aa2
--- /dev/null
+++ b/pythonetc/re_compile.md
@@ -0,0 +1,37 @@
+Always precompile regular expressions using `re.compile` if the expression is known in advance:
+
+```python
+# generate random string
+from string import printable
+from random import choice
+text = ''.join(choice(printable) for _ in range(10 * 8))
+
+# let's find numbers
+pat = r'\d(?:[\d\.]+\d)*'
+rex = re.compile(pat)
+
+%timeit re.findall(pat, text)
+# 2.08 µs ± 1.89 ns per loop
+
+# pre-compiled almost twice faster
+%timeit rex.findall(text)
+# 1.3 µs ± 68.8 ns per loop
+```
+
+The secret is that module-level `re` functions just compile the expression and call the corresponding method, no optimizations involved:
+
+```python
+def findall(pattern, string, flags=0):
+    return _compile(pattern, flags).findall(string)
+```
+
+If the expression is not known in advance but can be used repeatedly, consider using `functools.lru_cache`:
+
+```python
+from functools import lru_cache
+
+cached_compile = lru_cache(maxsize=64)(re.compile)
+
+def find_all(pattern, text):
+    return cached_compile(pattern).findall(text)
+```