diff --git a/backend/apps/agents/browser/browser_agent.py b/backend/apps/agents/browser/browser_agent.py
index fba6304b..64513aa8 100644
--- a/backend/apps/agents/browser/browser_agent.py
+++ b/backend/apps/agents/browser/browser_agent.py
@@ -2038,9 +2038,11 @@ async def run_browser_agent(
             "browser_id": browser_id,
             "summary": summary,
             # structured success signal the parent reads (replaces grepping the
-            # summary for a tag): true only if the run is honest AND, when the
-            # model called Done, it reported success. No-Done runs lean on honest.
-            "done": honest and (done_success if done_called else True),
+            # summary for a tag): true ONLY when the model explicitly finished via
+            # Done with success AND the honesty gate agreed. A run that just stopped
+            # (max turns, gave up, never called Done) is NOT a clean success, so the
+            # fast path recovers instead of shipping a silent half-finish.
+            "done": honest and done_called and done_success,
             # surface the honest failure to the parent so it doesn't treat a
             # did-nothing run as a success it can build on
             **({} if honest else {"error": summary}),
diff --git a/backend/apps/agents/browser/browser_loop.py b/backend/apps/agents/browser/browser_loop.py
index cc4a954b..b1335864 100644
--- a/backend/apps/agents/browser/browser_loop.py
+++ b/backend/apps/agents/browser/browser_loop.py
@@ -308,7 +308,8 @@ def replay_recheck_is_safe(action_log: list[dict]) -> bool:
 # never replay (the answer must be fresh), an action ask can.
 _INFO_ASK_RE = re.compile(
     r"\b(tell me|what(?:'s| is| are)|how (?:many|much)|count|list|summari[sz]e|"
-    r"extract|find out|read (?:me|the)|get the|give me|which|who (?:is|are)|report back)\b",
+    r"extract|find (?:me|out)|show me|look up|read (?:me|the)|get the|give me|which|"
+    r"who (?:is|are)|report back|most (?:viewed|popular|liked|recent|rated|watched)|top \d+)\b",
     re.I,
 )
 _ACTION_ASK_RE = re.compile(
diff --git a/backend/apps/agents/browser/browser_schema.py b/backend/apps/agents/browser/browser_schema.py
index 93d456ee..f8518321 100644
--- a/backend/apps/agents/browser/browser_schema.py
+++ b/backend/apps/agents/browser/browser_schema.py
@@ -816,7 +816,15 @@ SYSTEM_PROMPT = (
     "- To pull SPECIFIC FIELDS off a page (names + links from results, prices, table "
     "rows), call BrowserExtract with what you want; a helper model reads the page and "
     "hands you just the JSON. One call replaces BrowserGetText plus you reading 15k "
-    "chars, so prefer it whenever you know the fields you're after.\n\n"
+    "chars, so prefer it whenever you know the fields you're after.\n"
+    "- For ANY 'find me / list / get all / top N / most-viewed' gathering task, "
+    "BrowserExtract IS your tool: state the fields and let the helper read the rendered "
+    "page. Do NOT hand-roll CSS/DOM selectors in BrowserEvaluate to scrape a list, modern "
+    "sites (YouTube, LinkedIn, Amazon) obfuscate and lazy-render class names, so selector "
+    "scraping burns turns and silently returns the wrong nodes. If one BrowserEvaluate read "
+    "comes back empty or shaped wrong, STOP, that is your signal to switch to BrowserExtract, "
+    "not to debug another selector. The answer must end up IN your Done message, so gather it "
+    "for real; never report done on a gather task you couldn't actually read.\n\n"
 
     "## When you genuinely cannot proceed\n"
     "Use RequestHumanIntervention for:\n"
diff --git a/backend/tests/test_browser_agent_loop.py b/backend/tests/test_browser_agent_loop.py
index 27384085..505d5e3f 100644
--- a/backend/tests/test_browser_agent_loop.py
+++ b/backend/tests/test_browser_agent_loop.py
@@ -267,6 +267,21 @@ def test_done_tool_success_false_marks_not_done(monkeypatch):
     assert "login wall" in result["summary"]
 
 
+def test_run_that_never_calls_done_is_not_a_clean_success(monkeypatch):
+    # A run that does real work but stops with plain text (never calls Done) is a
+    # half-finish, not a clean success: done must be False so the fast path recovers
+    # instead of shipping a silent stop (the 'Task completed.' that wasn't).
+    BH._browser_history.clear(); BH._domain_notes.clear()
+    primary = FakeLLM([
+        Resp([_rp("click it"), _tu("BrowserClickIndex", index=3)]),
+        Resp([Blk("text", "I clicked the thing.")], stop_reason="end_turn"),
+    ])
+    aux = FakeAux()
+    _install(monkeypatch, primary, aux)
+    result = asyncio.run(BA.run_browser_agent(task="open the settings page", browser_id="b1", model="sonnet"))
+    assert result.get("done") is False  # no explicit Done -> not a clean success
+
+
 def test_early_perception_is_not_cut_short_before_any_action(monkeypatch):
     # Orienting on a cold/slow page can take several look-only turns; the stall
     # backstop must NOT fire before the agent has done anything (it only bounds a
@@ -1402,6 +1417,16 @@ def test_informational_gate_strips_outcome_boilerplate_on_tie_break():
     assert deliverable_is_informational(listy, "")
 
 
+def test_find_me_and_most_viewed_asks_are_informational():
+    from backend.apps.agents.browser.browser_loop import deliverable_is_informational
+    # the exact MKBHD task that previously slipped past the gate ("find me" + "most viewed")
+    assert deliverable_is_informational("", "find me his 50 most viewed vids")
+    assert deliverable_is_informational("", "show me the top 10 trending repos")
+    assert deliverable_is_informational("", "look up the cheapest flight")
+    # a pure action ask still is not informational
+    assert not deliverable_is_informational("", "send Tyler a message saying hi")
+
+
 def test_interstitial_dismiss_target_generalizable_and_safe():
     from backend.apps.agents.browser.browser_loop import interstitial_dismiss_target
     # a junk popup with a throwaway-dismiss control gets found (any site)