diff --git a/backend/apps/agents/browser/browser_agent.py b/backend/apps/agents/browser/browser_agent.py
index 57634785..fbf35b32 100644
--- a/backend/apps/agents/browser/browser_agent.py
+++ b/backend/apps/agents/browser/browser_agent.py
@@ -44,6 +44,7 @@ from backend.apps.agents.browser.browser_loop import (
     completion_is_honest,
     deliverable_is_informational,
     interstitial_dismiss_target,
+    is_publish_task,
     is_removal_task,
     recoverable_tool_error,
     replay_recheck_is_safe,
@@ -3069,8 +3070,17 @@ async def run_browser_agent(
         # turns while deleting nothing (measured live: repeat deletes all ghost-succeeded). Deletes
         # always run fresh through the delete-dispatch.
         p_is_removal = is_removal_task(skill_key_task)
-        logger.info(f"[browser-skills] record gate: honest={honest} informational={informational} removal={p_is_removal}")
-        if honest and not informational and not p_is_removal:
+        # The same trap as removals, one door down. A submit that never confirmed leaves an action_log
+        # of navigating and typing with NO send in it, so what distills is the scaffolding around a
+        # thing that did not happen. Measured live on reddit: the composer filled, then
+        # `send_button_found=False`, the agent blind-tapped a coordinate, nothing posted, and it still
+        # recorded a one-step "skill" (click the body textbox) for "create a text post and submit it".
+        # Replaying that reports done in one turn while posting nothing. A write only teaches us
+        # something once it actually went through.
+        p_unconfirmed_send = is_publish_task(skill_key_task) and not send_confirmed
+        logger.info(f"[browser-skills] record gate: honest={honest} informational={informational} "
+                    f"removal={p_is_removal} unconfirmed_send={p_unconfirmed_send}")
+        if honest and not informational and not p_is_removal and not p_unconfirmed_send:
             try:
                 rec_host = browser_skills.host_of(last_seen_url)
                 p_distilled = browser_skills.distill_steps(action_log)
diff --git a/backend/apps/agents/browser/browser_loop.py b/backend/apps/agents/browser/browser_loop.py
index 5050ffd7..c34b2250 100644
--- a/backend/apps/agents/browser/browser_loop.py
+++ b/backend/apps/agents/browser/browser_loop.py
@@ -298,6 +298,25 @@ def is_removal_task(task: str) -> bool:
     return bool(P_DELETE_INTENT_RE.search(task or ""))
 
 
+# Verbs that put something OUT into the world, as opposed to merely acting on a page. Deliberately
+# narrower than task_is_send, which only means "not an informational ask" and so counts a plain
+# "click the Search button": gating the skill store on THAT stopped the agent learning any click
+# task at all, which is the whole speed mechanism.
+P_PUBLISH_INTENT_RE = re.compile(
+    r"\b(post|submit|publish|send|tweet|comment|repl(?:y|ies)|dm|message)\b", re.I)
+
+
+def is_publish_task(task: str) -> bool:
+    """A task whose deliverable LEAVES the machine (a post, a reply, a message).
+
+    Keeps an unconfirmed write out of the skill store. Measured live on reddit: the composer filled,
+    `send_button_found=False`, the agent blind-tapped a coordinate, nothing posted, and a one-step
+    "skill" (click the body textbox) still got recorded for "create a text post and submit it".
+    Replaying that reports done in one turn while posting nothing, the same ghost the removal gate
+    already exists to stop."""
+    return bool(P_PUBLISH_INTENT_RE.search(task or ""))
+
+
 def deliverable_is_informational(summary: str, task: str = "") -> bool:
     """True if the run's final answer is GATHERED CONTENT (a list/report the model
     extracted or judged), not a short action confirmation. A deterministic replay
diff --git a/backend/tests/test_browser_agent_loop.py b/backend/tests/test_browser_agent_loop.py
index fed80478..2ac3bf5c 100644
--- a/backend/tests/test_browser_agent_loop.py
+++ b/backend/tests/test_browser_agent_loop.py
@@ -1913,3 +1913,47 @@ def test_login_wall_skip_does_not_remember(monkeypatch, tmp_path):
     p_run_settled(task="log into acme", browser_id="b1", model="sonnet")
 
     assert not H.is_authenticated("acme.example")
+
+
+def test_an_unconfirmed_post_teaches_us_nothing(monkeypatch):
+    """Measured live on reddit 2026-07-31: the composer filled, then `send_button_found=False`, the
+    agent blind-tapped a coordinate, NOTHING posted, and it still recorded a one-step "skill" (click
+    the body textbox) for "create a text post and submit it". Replaying that reports done in one turn
+    while posting nothing, the ghost class the removal gate already exists to stop.
+
+    Uses the SAME tool shape as test_skill_is_recorded_then_replayed_with_zero_llm_calls, which does
+    record, so the only difference here is that the ask publishes and the send never confirmed."""
+    import backend.apps.agents.browser.browser_skills as SK
+    SK.clear()
+    BH.BROWSER_HISTORY.clear(); BH.DOMAIN_NOTES.clear()
+    task = "submit the post"
+    primary = FakeLLM([
+        Resp([p_rp("find the control"), p_tu("BrowserListInteractives")]),
+        Resp([p_rp("click it"), p_tu("BrowserClickIndex", index=1)]),
+        Resp([Blk("text", "Done, I sent it for you.")], stop_reason="end_turn"),
+    ])
+    p_install(monkeypatch, primary, FakeAux())
+    asyncio.run(BA.run_browser_agent(
+        task=task, browser_id="b1", model="sonnet", initial_url=DOC_URL,
+    ))
+    assert SK.find_skill("docs.google.com", task) is None, (
+        "an unconfirmed publish must never be distilled into a replayable skill")
+
+
+def test_a_plain_click_task_still_learns(monkeypatch):
+    """The publish gate must stay narrow. `task_is_send` only means "not an informational ask", so
+    keying on it stopped the agent learning ANY click task, which is the whole speed mechanism."""
+    import backend.apps.agents.browser.browser_skills as SK
+    SK.clear()
+    BH.BROWSER_HISTORY.clear(); BH.DOMAIN_NOTES.clear()
+    primary = FakeLLM([
+        Resp([p_rp("find it"), p_tu("BrowserListInteractives")]),
+        Resp([p_rp("click it"), p_tu("BrowserClickIndex", index=1)]),
+        Resp([Blk("text", "Done, clicked Search.")], stop_reason="end_turn"),
+    ])
+    p_install(monkeypatch, primary, FakeAux())
+    asyncio.run(BA.run_browser_agent(
+        task="click the Search button", browser_id="b1", model="sonnet", initial_url=DOC_URL,
+    ))
+    assert SK.find_skill("docs.google.com", "click the Search button") is not None, (
+        "a non-publishing action task must still be learnable")