diff --git a/core/archipelago/src/assistant/mod.rs b/core/archipelago/src/assistant/mod.rs
index 7b32dadc..d241f8e8 100644
--- a/core/archipelago/src/assistant/mod.rs
+++ b/core/archipelago/src/assistant/mod.rs
@@ -771,18 +771,31 @@ instead of fabricating a tool call.\n\n\
 That refusal rule is about ACTIONS ON THIS NODE, not about conversation. You are also the \
 owner's general assistant: answer ordinary questions, explain things, and give recommendations \
 (films, music, podcasts, books, reading) from your own knowledge, exactly as a capable \
-assistant would. Those answers are what the content surfaces in this app render as cards, so \
-declining to answer them leaves the owner staring at an empty panel. Only a request to CHANGE \
-or READ something on the node needs a tool — and if no tool covers it, say so plainly rather \
-than pretending. Never let a general question be refused merely because no tool matches it.\n\n\
+assistant would. Only a request to CHANGE or READ something on the node needs a tool — and if \
+no tool covers it, say so plainly rather than pretending. Never let a general question be \
+refused merely because no tool matches it.\n\n\
 When the question is about what actually EXISTS here — this node's own shared files, a peer's \
 catalogue, purchased items, the film catalogue — call the content tool rather than answering \
 from memory or from earlier in this conversation, and call it once per place the question spans \
-(asking about \"films\" covers both the catalogue and the peers). The app renders whatever that \
-tool returns as cards beside your reply, so your text should introduce and summarise the result \
-— how many, anything notable — rather than re-listing every title, size and price in prose. A \
-long enumeration duplicates the cards the operator is already looking at. If the tool comes \
-back empty, say so plainly; do not fill the gap with remembered or invented items.";
+(asking about \"films\" covers both the catalogue and the peers). The same applies to \
+RECOMMENDATIONS of a kind this node could hold: \"recommend me ten sci-fi films\" gets a \
+catalogue-and-peers check FIRST, then your own picks on top. And never end a turn by offering \
+to check (\"would you like me to look?\") — looking costs one tool call; do it, then answer \
+with what you found. The app renders whatever the tool returns as cards beside your reply, so \
+your text should introduce and summarise the result — how many, anything notable — rather than \
+re-listing every title, size and price in prose. If the tool comes back empty, say so plainly; \
+do not present remembered items as if they were found here.\n\n\
+Recommendations from your own knowledge are still wanted, and the app renders each one as a \
+rich preview card when you tag it. Put the tag on the SAME line as the title's one-line \
+reason, in EXACTLY these formats:\n\
+- [[film_ext:Title|Year|Director]] for a film, e.g. [[film_ext:Arrival|2016|Denis Villeneuve]]\n\
+- [[tv_ext:Title|Year|Creator]] for a TV series\n\
+- [[song_ext:Title|Artist|Year]] for a song\n\
+- [[book_ext:Title|Author|Year]] for a book\n\
+- [[podcast_ext:Title|Host|Year]] for a podcast\n\
+Tag only specific, real titles you are confident about — never invent a title to fill a tag, \
+and never tag a title the tool already returned (its card is already on screen). An untagged \
+recommendation renders as plain text and is easily missed.";
 
 pub fn build_system_prompt(visible_tools: &[tools::ToolDef]) -> String {
     let mut prompt = String::from(SYSTEM_PROMPT_PREAMBLE);
@@ -1228,6 +1241,35 @@ mod tests {
         );
     }
 
+    /// Operator report 2026-08-07: "recommend me 10 scifi films" answered
+    /// with a prose-only knowledge list and an OFFER to check the
+    /// catalogue — the model classified a recommendation as pure
+    /// conversation, never called the content tool, and the reply carried
+    /// none of the preview tags the iframe renders as cards. The preamble
+    /// must pin both behaviors: check-first on node-holdable
+    /// recommendations, and the exact tag formats for knowledge picks.
+    #[test]
+    fn system_prompt_teaches_discovery_first_and_preview_tags() {
+        let prompt = build_system_prompt(&[]);
+        assert!(
+            prompt.contains("catalogue-and-peers check FIRST"),
+            "prompt must order a catalogue/peers check before freestyle recommendations"
+        );
+        assert!(
+            prompt.contains("never end a turn by offering to check"),
+            "prompt must ban the 'would you like me to look?' stall"
+        );
+        for tag in [
+            "[[film_ext:Title|Year|Director]]",
+            "[[tv_ext:Title|Year|Creator]]",
+            "[[song_ext:Title|Artist|Year]]",
+            "[[book_ext:Title|Author|Year]]",
+            "[[podcast_ext:Title|Host|Year]]",
+        ] {
+            assert!(prompt.contains(tag), "prompt must teach {tag}");
+        }
+    }
+
     /// D-16: revoking a category takes effect on the very next
     /// `granted_categories` resolution — not only on the next session /
     /// process restart.