diff --git a/docs/DEVELOPER_NOTES.md b/docs/DEVELOPER_NOTES.md
new file mode 100644
index 0000000..019e26d
--- /dev/null
+++ b/docs/DEVELOPER_NOTES.md
@@ -0,0 +1,74 @@
+# Developer Notes: Documentation Architecture
+
+This document explains the documentation build system and why it's set up the way it is.
+
+## The Recursion Problem
+
+The documentation for `markdown-code-runner` uses `markdown-code-runner` itself to generate content. This creates a potential recursion problem:
+
+1. The docs contain **examples** showing how to use the tool (CODE:START, OUTPUT:START markers, etc.)
+2. If these examples were executed by the tool, their OUTPUT sections would be filled with actual output
+3. This would make them useless as examples - users need to see the *template* syntax, not executed results
+
+## How We Solve It
+
+### 1. CODE:SKIP Directive
+
+Use `<!-- CODE:SKIP -->` before example code blocks to prevent them from being executed:
+
+```markdown
+<!-- CODE:SKIP --> <!-- Prevents the next code block from running -->
+```python
+print("This is an example that won't be executed")
+```
+```
+
+### 2. PLACEHOLDER Pattern for Templates
+
+OUTPUT sections in the repository contain only placeholder text:
+
+```markdown
+<!-- OUTPUT:START -->
+<!-- PLACEHOLDER --> Output is generated during CI build. We don't commit generated content to keep docs copyable and avoid recursion. See docs/docs_gen.py
+<!-- OUTPUT:END -->
+```
+
+This placeholder gets replaced with actual content during CI builds, but the generated content is **never committed** back to the repository.
+
+### 3. Nested HTML Comment Escaping
+
+When writing code that references markdown-code-runner markers inside HTML comments, you must escape them to avoid breaking markdown parsing:
+
+```python
+# BAD - nested HTML comments break parsing:
+<!-- content = re.sub(r"<!-- SECTION:START -->.*?<!-- SECTION:END -->", ...) -->
+
+# GOOD - use string concatenation to escape:
+<!-- start_marker = "<!-" + "- SECTION:START -" + "->" -->
+<!-- end_marker = "<!-" + "- SECTION:END -" + "->" -->
+<!-- content = re.sub(start_marker + ".*?" + end_marker, ...) -->
+```
+
+## Documentation Build Process
+
+1. **Source files** in `docs/` contain CODE blocks that pull content from README.md
+2. **CI runs** `docs/docs_gen.py` which executes `markdown-code-runner` on all doc files
+3. **Generated output** is used to build the documentation site
+4. **Generated content is NOT committed** - only the templates with PLACEHOLDERs exist in the repo
+
+## File Structure
+
+- `docs/docs_gen.py` - Script that processes all markdown files for documentation build
+- `docs/*.md` - Documentation templates with CODE/OUTPUT blocks
+- `README.md` - Source of truth for content (uses SECTION markers)
+
+## Pre-commit Hook
+
+The `Verify docs templates have placeholders` pre-commit hook ensures that OUTPUT sections contain PLACEHOLDER text, preventing accidental commits of generated content.
+
+## Why Not Commit Generated Output?
+
+1. **Copyability**: Users copying examples from the docs get clean templates, not filled-in output
+2. **Recursion avoidance**: Examples remain as examples, not executed results
+3. **Single source of truth**: README.md is the source; docs pull from it
+4. **Cleaner diffs**: No noise from regenerated output in every commit