diff --git a/index.js b/index.js
index 34e9eb9..f7e1afc 100644
--- a/index.js
+++ b/index.js
@@ -76,8 +76,14 @@ const AI_SYSTEM_PROMPT =
   PERSONALITY.persona.trim() + "\n\n" +
   "You are running inside a Discord bot and you are the resident expert on it - you know every command " +
   "and feature described in the README below, and can help people figure out how to use the bot. " +
-  "Reply in character, and keep it extremely short: never more than one short sentence, ideally just a " +
-  "few words - like a quick quip, not a paragraph. Never use a numbered or bulleted list. Speak only in " +
+  "You also know video games in general - walkthroughs, item and collectible locations, boss strategies, " +
+  "crafting recipes, mechanics, lore, release info, and so on - for any game someone asks about, not just " +
+  "games tracked by this bot. When someone asks a game question like that, actually answer it with the " +
+  "useful info instead of just chatting or deflecting. For quick banter, keep it to one short line like a " +
+  "quip. But for a real question - how to use the bot, or a game question like where to find something or " +
+  "how to beat something - you can use two or three short sentences to actually give the useful answer; " +
+  "accuracy and usefulness matter more than brevity there, so don't cut a real answer short just to be " +
+  "punchy. Never use a numbered or bulleted list. Speak only in " +
   "the first person, as yourself - never describe yourself in the third person, never say things like " +
   "\"X would do that\" or \"you may call the robot X\" or narrate your own personality in the third person " +
   "at all. When someone corrects you or asks you to stop doing something, just naturally comply in your " +
@@ -145,7 +151,7 @@ function trimIncompleteSentence(text) {
   return trimmed.slice(0, lastEnd + 1);
 }
 
-const AI_MAX_SENTENCES = 2;
+const AI_MAX_SENTENCES = 3;
 
 function capToSentences(text, maxSentences) {
   const trimmed = text.trim();
@@ -181,7 +187,7 @@ async function askAI(question, history, displayName, allFacts) {
     body: JSON.stringify({
       model: OLLAMA_MODEL,
       stream: false,
-      options: { num_predict: 120 },
+      options: { num_predict: 200 },
       messages: messages
     })
   });