diff --git a/README.md b/README.md
index 0e846dc..9909d26 100644
--- a/README.md
+++ b/README.md
@@ -21,15 +21,13 @@ programmatic/LLM consumption.
 
 ## Architecture
 
-```
-tokenize(text)          generic S-expression tokenizer
-parse_sexp(text)        builds { tag, items[] } tree
-        │
-        ▼
-run_query(ext, query, root, args)   file-type + query dispatch → plain data object
-        │
-        ├── render_human(ext, query, data)   formatted tables for terminal
-        └── render_json(data)                JSON.stringify for LLM / scripts
+```mermaid
+flowchart TD
+    A["tokenize(text)"] --> B["parse_sexp(text)\n{ tag, items[] } tree"]
+    B --> C["run_query(ext, query, root, args)\nfile-type + query dispatch"]
+    C --> D["plain data object"]
+    D --> E["render_human(ext, query, data)\nformatted tables for terminal"]
+    D --> F["render_json(data)\nJSON.stringify for LLM / scripts"]
 ```
 
 Query functions are pure data transformers — no I/O, no format awareness.