diff --git a/rig-core/examples/gemini_agent.rs b/rig-core/examples/gemini_agent.rs
index f221886a..978a7cff 100644
--- a/rig-core/examples/gemini_agent.rs
+++ b/rig-core/examples/gemini_agent.rs
@@ -12,23 +12,22 @@ async fn main() -> Result<(), anyhow::Error> {
     // Create agent with a single context prompt
     let agent = client
         .agent(gemini::completion::GEMINI_1_5_PRO)
-        .preamble("Be precise and concise.")
+        .preamble("Be creative and concise. Answer directly and clearly.")
         .temperature(0.5)
         .max_tokens(8192)
-        .additional_params(
-            serde_json::to_value(GenerationConfig {
-                top_k: Some(1),
-                top_p: Some(0.95),
-                candidate_count: Some(1),
-                ..Default::default()
-            })
-            .unwrap(),
-        ) // Unwrap the Result to get the Value
+        .additional_params(serde_json::to_value(GenerationConfig {
+            top_k: Some(1),
+            top_p: Some(0.95),
+            candidate_count: Some(1),
+            ..Default::default()
+        })?) // Unwrap the Result to get the Value
         .build();
 
+    tracing::info!("Prompting the agent...");
+
     // Prompt the agent and print the response
     let response = agent
-        .prompt("How much wood would a woodchuck chuck if a woodchuck could chuck wood?")
+        .prompt("How much wood would a woodchuck chuck if a woodchuck could chuck wood? Infer an answer.")
         .await?;
     println!("{}", response);
 
diff --git a/rig-core/src/providers/gemini/completion.rs b/rig-core/src/providers/gemini/completion.rs
index 12cd87be..c535a56c 100644
--- a/rig-core/src/providers/gemini/completion.rs
+++ b/rig-core/src/providers/gemini/completion.rs
@@ -36,7 +36,6 @@ pub struct GenerateContentResponse {
     pub usage_metadata: Option<UsageMetadata>,
 }
 
-// Define the struct for a Candidate
 #[derive(Debug, Deserialize)]
 #[serde(rename_all = "camelCase")]
 pub struct ContentCandidate {
@@ -341,7 +340,7 @@ pub struct GenerateContentRequest {
     pub safety_settings: Option<Vec<SafetySetting>>,
     /// Optional. Developer set system instruction(s). Currently, text only.
     /// https://ai.google.dev/gemini-api/docs/system-instructions?lang=rest
-    pub system_instruction: Option<String>,
+    pub system_instruction: Option<Content>,
     // cachedContent: Optional<String>
 }
 
@@ -414,17 +413,7 @@ impl completion::CompletionModel for CompletionModel {
         &self,
         mut completion_request: CompletionRequest,
     ) -> Result<completion::CompletionResponse<GenerateContentResponse>, CompletionError> {
-        // QUESTION: Why do Anthropic/openAi implementation differ here? OpenAI adds the preamble but Anthropic does not.
-
-        let mut full_history = if let Some(preamble) = &completion_request.preamble {
-            vec![completion::Message {
-                role: "system".into(),
-                content: preamble.clone(),
-            }]
-        } else {
-            vec![]
-        };
-
+        let mut full_history = Vec::new();
         full_history.append(&mut completion_request.chat_history);
 
         let prompt_with_context = completion_request.prompt_with_context();
@@ -471,9 +460,16 @@ impl completion::CompletionModel for CompletionModel {
                     .collect(),
             ),
             tool_config: None,
-            system_instruction: None,
+            system_instruction: Some(Content {
+                parts: vec![Part {
+                    text: "system".to_string(),
+                }],
+                role: Some("system".to_string()),
+            }),
         };
 
+        tracing::info!("Request: {:?}", request);
+
         let response = self
             .client
             .post(&format!("/v1beta/models/{}:generateContent", self.model))