diff --git a/src/context_system/prompt_assembly.py b/src/context_system/prompt_assembly.py
index 41bb12ae..066cf6f8 100644
--- a/src/context_system/prompt_assembly.py
+++ b/src/context_system/prompt_assembly.py
@@ -1002,12 +1002,44 @@ def _emit_group(
     "understand and review your work. This is CRITICAL.\n"
     "  - To read files use Read instead of cat, head, tail, or sed\n"
     "  - To edit files use Edit instead of sed or awk\n"
-    "  - To create files use Write instead of cat with heredoc or echo\n"
+    # The scope clause is the port's own rationale ("allows the user to
+    # better understand and review your work") applied consistently: it is
+    # about files the user reviews. Unscoped, the rule also caught scratch
+    # scripts, which cost a step every time — clawcodex authors a probe
+    # script with Write and then spends a second step running it, where the
+    # current Claude Code does both in one Bash call. Measured on seven
+    # shared terminal-bench 2.1 tasks at the same model and effort
+    # (2026-07-26): Claude Code authors 23.4% of its Bash calls via heredoc
+    # against clawcodex's 2.7%, and spends 13.6% of its tool calls on
+    # Write/Edit against clawcodex's 25.5%. The reference's own wording is a
+    # single soft line ("Prefer the dedicated file/search tools over shell
+    # commands when one fits"), so the absolute reading is not faithful either.
+    # Location is deliberately unspecified: it makes no permission difference.
+    # check_accept_edits_bash (src/permissions/bash_mode_validation.py:304)
+    # gates on _has_shell_redirection, not on where the redirect points --
+    # `cat` is in ACCEPT_EDITS_READ_ONLY_COMMANDS and that branch requires
+    # `not redirection_anywhere` -- so every authoring form returns False and
+    # prompts, in-roots and /tmp alike (verified for `cat > /app/x <<PY`,
+    # `cat > /tmp/x <<PY`, `cat > ./x <<PY`, `printf 'x' > /app/x`).
+    #
+    # KNOWN COST, accepted deliberately: in acceptEdits mode Write is
+    # auto-accepted while a Bash heredoc prompts, so taking this option trades
+    # a saved step for an approval prompt there. It is a clear win only where
+    # Bash is already unprompted (headless, bypass). Hence "may" rather than
+    # an instruction -- the model still has Write available and should use it
+    # when a prompt would cost more than the step saves.
+    "  - To create files use Write instead of cat with heredoc or echo. A "
+    "throwaway script you run once and discard is part of the command "
+    "rather than a deliverable, and may instead be authored inline in the "
+    "same Bash call that runs it.\n"
     "  - To search for files use Glob instead of find or ls\n"
     "  - To search the content of files, use Grep instead of grep or rg\n"
-    "  - Reserve using the Bash exclusively for system commands and terminal "
-    "operations that require shell execution. If you are unsure and there "
-    "is a relevant dedicated tool, default to using the dedicated tool.\n"
+    # "exclusively" was dropped: it cannot coexist with the throwaway-script
+    # carve-out above, and a prompt that states a rule and its exception in
+    # absolute terms leaves the model to guess which one governs.
+    "  - Reserve Bash for system commands and terminal operations that "
+    "require shell execution. If you are unsure and there is a relevant "
+    "dedicated tool, default to using the dedicated tool.\n"
     "- Break down and manage your work with the TaskCreate tool.\n"
     # Two clauses were dropped in the port, and together they are what turns
     # a permission into a practice: the imperative to maximize, and the
diff --git a/tests/test_system_prompt_full.py b/tests/test_system_prompt_full.py
index 45af95b7..c41b1beb 100644
--- a/tests/test_system_prompt_full.py
+++ b/tests/test_system_prompt_full.py
@@ -161,6 +161,38 @@ def test_tools_prompt_pushes_parallel_tool_calls(self):
             "the dependency carve-out is what makes parallelizing safe to act on"
         )
 
+    def test_throwaway_scripts_may_be_authored_inline(self):
+        """A one-shot script is part of the command, not a deliverable.
+
+        The Write preference must survive for deliverables — that is what
+        keeps real edits reviewable as diffs — while a throwaway script the
+        agent runs once may be authored inside the Bash call that runs it,
+        saving a step.
+
+        Stated as permission ("may"), not instruction, because the tradeoff
+        is context-dependent: in acceptEdits mode Write is auto-accepted
+        while any Bash redirect prompts (check_accept_edits_bash gates on
+        _has_shell_redirection, so the target path is irrelevant), so the
+        saved step costs an approval there.
+
+        No location is specified — /tmp and in-roots paths prompt
+        identically, so mandating one buys nothing.
+        """
+        prompt = build_full_system_prompt(use_cache=False)
+        assert "To create files use Write instead of cat with heredoc" in prompt, (
+            "the deliverable path must still prefer Write, for reviewable diffs"
+        )
+        assert "throwaway script you run once and discard" in prompt
+        assert "may instead be authored inline" in prompt, (
+            "permission, not instruction — the tradeoff depends on whether "
+            "Bash is already unprompted in the caller's permission mode"
+        )
+        tools_section = prompt.split("# Using your tools")[1].split("\n- Break")[0]
+        assert "/tmp" not in tools_section
+        assert "exclusively" not in tools_section, (
+            "an absolute 'Reserve Bash exclusively' contradicts the carve-out"
+        )
+
     def test_identity_prompt_backward_compat(self):
         """_IDENTITY_PROMPT is an alias for _INTRO_SECTION."""
         assert _IDENTITY_PROMPT is _INTRO_SECTION