diff --git a/DEVELOPMENT.md b/DEVELOPMENT.md
new file mode 100644
index 0000000..848d0c2
--- /dev/null
+++ b/DEVELOPMENT.md
@@ -0,0 +1,75 @@
+# [DRAFT] Transformers
+
+You can inject modules into Universql at start time, by passing `--module-paths ./my_transformers`.
+
+Here is the abstract class for transformer:
+
+```python
+class Transformer:
+    def __init__(self,
+                 # allows us to call this transformer when base catalog is snowflake
+                 source_engine: SnowflakeCatalog,
+                 # if this is generic as Catalog, can automatically invoke
+                 # .transform() no matter where it's running on
+                 target_engine: Catalog
+                 ):
+        self.source_engine = source_engine
+        self.target_engine = target_engine
+
+    def transform_sql(self, expression: Expression) -> Expression:
+        return expression
+
+    def transform_result(self, response: Response):
+        return response
+
+    def transform_request(self, request: Request):
+        return request
+```
+
+Here is an example:
+
+```python
+# Rewrites Snowflake timestamp time to DuckDB
+class FixTimestampTypes(Transformer):
+    def __init__(self, source_engine: SnowflakeCatalog, target_engine: DuckDBCatalog):
+        super().__init__(source_engine, target_engine)
+        
+    def transform_sql(self, expression):
+        if isinstance(expression, sqlglot.exp.DataType):
+            if expression.this.value in ["TIMESTAMPLTZ", "TIMESTAMPTZ"]:
+                return sqlglot.exp.DataType.build("TIMESTAMPTZ")
+            if expression.this.value in ["VARIANT"]:
+                return sqlglot.exp.DataType.build("JSON")
+
+        return expression
+```
+
+The engine will look into the `__init__` and call the transformer only when the catalogs type match.
+
+Things we can move to transformers:
+
+```python
+class RewriteCreateAsIceberg(Transformer):
+
+    def transform_sql(self, expression: Expression) -> Expression:
+        # re-write CREATE TABLE as CREATE ICEBERG TABLE
+        return expression
+```
+
+For stage integration, something like:
+
+```python
+class SnowflakeStageTransformer(Transformer):
+    def __init__(self, source_engine: SnowflakeCatalog, target_engine: DuckDBCatalog):
+        super().__init__(source_engine, target_engine)
+        
+    def transform_sql(self, ast: Expression) -> Expression:
+        if isinstance(ast, sqlglot.exp.Var) and ast.name.startswith('@'):
+            # transform into full path and create secret on duckdb
+            self.target_engine.duckdb.sql("select from stage information_schema.stages where ..")
+            if not_exists:
+                self.source_engine.executor().execute_raw("get stage info from fs")
+                self.target_engine.duckdb.sql("INSERT INTO information_schema.stages ...")
+            return new_ast_with_full_path
+        return ast
+```
\ No newline at end of file