diff --git a/crates/tinyagents-graph/src/todos/test.rs b/crates/tinyagents-graph/src/todos/test.rs
index dc7a5505..a3a0efdd 100644
--- a/crates/tinyagents-graph/src/todos/test.rs
+++ b/crates/tinyagents-graph/src/todos/test.rs
@@ -355,6 +355,33 @@ mod tool_tests {
             .expect("successful todo result has a JSON payload")
     }
 
+    /// The description is what makes a model treat the list as bookkeeping
+    /// rather than as the work itself: without the "same response carries the
+    /// next call" rule, models write a list, stop, and wait to be prompted;
+    /// without the "only after its work has run" rule they tick items off
+    /// ahead of doing them.
+    #[test]
+    fn description_states_when_an_item_may_be_completed_and_that_writing_is_not_working() {
+        let tool = TodoTool::new(store());
+        let description = Tool::description(&tool);
+        assert!(
+            description.contains("only after its work has actually run"),
+            "an item is completed after its result exists: {description}"
+        );
+        assert!(
+            description.contains("bookkeeping, not work"),
+            "writing the list is not the work: {description}"
+        );
+        assert!(
+            description.contains("immediately carry out the next step"),
+            "the model advances after bookkeeping: {description}"
+        );
+        assert!(
+            description.contains("next model turn"),
+            "sequential providers may advance on their next turn: {description}"
+        );
+    }
+
     #[test]
     fn todo_tools_builds_a_single_tool() {
         let tools = todo_tools(store());
diff --git a/crates/tinyagents-graph/src/todos/tool.rs b/crates/tinyagents-graph/src/todos/tool.rs
index f84e9a2f..dcabec4e 100644
--- a/crates/tinyagents-graph/src/todos/tool.rs
+++ b/crates/tinyagents-graph/src/todos/tool.rs
@@ -25,9 +25,12 @@ const TODO_TOOL_NAME: &str = "todo";
 
 const TODO_DESCRIPTION: &str = "Your todo list for this thread. Pass the complete list every \
     time; it replaces what was there. Use it for work with 3+ steps: write the steps up front, \
-    keep exactly one `in_progress`, mark each `completed` the moment it is done. Omit `todos` \
-    to read the current list. The list is bound automatically to the current thread — do not \
-    pass a thread id.";
+    keep exactly one `in_progress`, mark each `completed` only after its work has actually run \
+    and its result is in this conversation. Writing the list is bookkeeping, not work: after \
+    updating it, immediately carry out the next step. Providers that cannot issue parallel tool \
+    calls may make that call in the next model turn, and one update per response is enough. Omit \
+    `todos` to read the current list. The list is bound \
+    automatically to the current thread — do not pass a thread id.";
 
 /// The `todo` harness [`Tool`], backed by a [`Store`](tinyagents_harness::store::Store).
 pub struct TodoTool {