diff --git a/public/src/chat.js b/public/src/chat.js
index e2e37f9..dfd869e 100644
--- a/public/src/chat.js
+++ b/public/src/chat.js
@@ -5,9 +5,9 @@ import { API_KEY, getAPIConfig } from "./config.js";
 import { getSelectedModel, getSelectedModelEndpoint, getSelectedModelMaxTokens, getSelectedModelMaxCompletionTokens } from "./models.js";
 import { groqAwareStreamingFetch } from "./groq-fetch.js";
 
-// Rough token estimation (3 chars = 1 token to be conservative)
+// Rough token estimation (4 chars = 1 token average for English)
 function estimateTokens(text) {
-	return Math.ceil(text.length / 3);
+	return Math.ceil(text.length / 4);
 }
 
 // Calculate available output tokens
diff --git a/public/src/storage.js b/public/src/storage.js
index 292741a..f51f089 100644
--- a/public/src/storage.js
+++ b/public/src/storage.js
@@ -34,21 +34,31 @@ export function loadConversationHistory() {
 	return [];
 }
 
+// Get just the custom prompt for attention echo (not full system message with page content)
+function getCustomPromptEcho() {
+	if (typeof window !== "undefined" && window.UNCLOSEAI_SYSTEM_PROMPT) {
+		return window.UNCLOSEAI_SYSTEM_PROMPT;
+	}
+	return null;
+}
+
 export async function clearConversationHistory() {
 	localStorage.removeItem(getPageSpecificKey("hermes-conversation-history"));
 	const systemContent = await getSystemMessage();
-	return [
-		{ role: "system", content: systemContent },
-		{ role: "assistant", content: systemContent },
-	];
+	const echo = getCustomPromptEcho();
+	const history = [{ role: "system", content: systemContent }];
+	if (echo) history.push({ role: "assistant", content: echo });
+	return history;
 }
 
 // Initialize chat history
 export async function initializeChatHistory() {
 	const systemContent = await getSystemMessage();
-	return [
+	const echo = getCustomPromptEcho();
+	const history = [
 		{ role: "system", content: systemContent },
 		...loadConversationHistory(),
-		{ role: "assistant", content: systemContent },
 	];
+	if (echo) history.push({ role: "assistant", content: echo });
+	return history;
 }
diff --git a/public/src/translation.js b/public/src/translation.js
index 1410176..58efa0b 100644
--- a/public/src/translation.js
+++ b/public/src/translation.js
@@ -75,7 +75,7 @@ async function* sendMessageWithHistory(messageHistory) {
 	// Calculate available tokens for output based on current prompt
 	const modelMaxTokens = getSelectedModelMaxTokens();
 	const inputText = messageHistory.map(msg => msg.content).join('');
-	const inputTokens = Math.ceil(inputText.length / 3); // Conservative token estimation
+	const inputTokens = Math.ceil(inputText.length / 4); // Simple token estimation
 	const outputReserve = Math.floor(modelMaxTokens * 0.4); // Reserve 40% for output
 	let availableTokens = Math.max(512, modelMaxTokens - inputTokens - outputReserve);
 	
diff --git a/public/src/tts.js b/public/src/tts.js
index 1c42518..0bdb657 100644
--- a/public/src/tts.js
+++ b/public/src/tts.js
@@ -143,7 +143,7 @@ Your task:
 7. Start with the article title, then the main content
 
 Return only the article text ready for TTS, nothing else.`;
-	const inputTokens = Math.ceil((systemPrompt + preCleanedContent).length / 3);
+	const inputTokens = Math.ceil((systemPrompt + preCleanedContent).length / 4);
 	const buffer = Math.max(2048, Math.floor(inputTokens * 1.5)); // Use 1.5x input tokens as buffer, minimum 2048
 	const calculatedAvailable = Math.max(100, modelMaxTokens - inputTokens - buffer);
 	
diff --git a/public/src/ui.js b/public/src/ui.js
index c1a02ed..b2ba880 100644
--- a/public/src/ui.js
+++ b/public/src/ui.js
@@ -141,9 +141,9 @@ async function* sendMessageWithCustomHistory(messageHistory) {
 	const { getSelectedModelMaxTokens, getSelectedModelMaxCompletionTokens } = await import("./models.js");
 	const modelMaxTokens = getSelectedModelMaxTokens();
 	
-	// Estimate input tokens for intro generation (3 chars = 1 token to be conservative)
+	// Estimate input tokens for intro generation
 	const inputText = messageHistory.map(msg => msg.content).join('');
-	const inputTokens = Math.ceil(inputText.length / 3);
+	const inputTokens = Math.ceil(inputText.length / 4);
 	
 	// For intro generation, be more generous - we want complete intros
 	// With large context models, we have plenty of room
diff --git a/public/src/uncloseai-embed-modal.js b/public/src/uncloseai-embed-modal.js
index 5cf9f36..67b9dfb 100644
--- a/public/src/uncloseai-embed-modal.js
+++ b/public/src/uncloseai-embed-modal.js
@@ -1499,10 +1499,11 @@ You have complete knowledge of this page content and can reference any details,
 
 				// Create new history with updated system message and loaded conversation
 				const systemContent = await getSystemMessage();
+				const echo = (typeof window !== "undefined" && window.UNCLOSEAI_SYSTEM_PROMPT) ? window.UNCLOSEAI_SYSTEM_PROMPT : null;
 				const newHistory = [
 					{ role: "system", content: systemContent },
 					...history,
-					{ role: "assistant", content: systemContent },
+					...(echo ? [{ role: "assistant", content: echo }] : []),
 				];
 
 				// Update the chat.js module's history
@@ -1648,10 +1649,11 @@ You have complete knowledge of this page content and can reference any details,
 								const { updateChatHistory } = await import("./chat.js");
 								const { getSystemMessage } = await import("./config.js");
 								const systemContent = await getSystemMessage();
+								const echo = (typeof window !== "undefined" && window.UNCLOSEAI_SYSTEM_PROMPT) ? window.UNCLOSEAI_SYSTEM_PROMPT : null;
 								const newHistory = [
 									{ role: "system", content: systemContent },
 									...currentHistory,
-									{ role: "assistant", content: systemContent },
+									...(echo ? [{ role: "assistant", content: echo }] : []),
 								];
 								await updateChatHistory(newHistory);
 							}
@@ -1967,10 +1969,11 @@ You have complete knowledge of this page content and can reference any details,
 
 			// Create new history with system message, intro, and echo for attention
 			const systemContent = await getSystemMessage();
+			const echo = (typeof window !== "undefined" && window.UNCLOSEAI_SYSTEM_PROMPT) ? window.UNCLOSEAI_SYSTEM_PROMPT : null;
 			const newHistory = [
 				{ role: "system", content: systemContent },
 				{ role: "assistant", content: response },
-				{ role: "assistant", content: systemContent },
+				...(echo ? [{ role: "assistant", content: echo }] : []),
 			];
 
 			// Update the chat.js module's history