diff --git a/django/aiservice/core/languages/js_ts/prompts/react_testgen/system_prompt.md b/django/aiservice/core/languages/js_ts/prompts/react_testgen/system_prompt.md
index 74a872d92..9d5f0e36e 100644
--- a/django/aiservice/core/languages/js_ts/prompts/react_testgen/system_prompt.md
+++ b/django/aiservice/core/languages/js_ts/prompts/react_testgen/system_prompt.md
@@ -111,6 +111,15 @@ If the component has no buttons, inputs, or interactive elements, you MUST still
 - Example: `const { rerender } = render(<Display data={data} />); rerender(<Display data={data} />); rerender(<Display data={newData} />);`
 - This produces update-phase markers that Codeflash needs for A/B comparison.
 
+**Assertion Safety Rules (Reduce False Failures)**
+- Do NOT assert specific `aria-label` values unless they are explicitly visible in the source code.
+- Do NOT assert element counts (`getAllByRole(...).toHaveLength(N)`) unless the count is deterministic from props.
+- Prefer `screen.getByText(...)` for visible text and `screen.getByRole('button')` for interactive elements.
+- When asserting DOM structure, use `toBeInTheDocument()` instead of specific positions or counts.
+- For components with dynamic data (lists, tables), assert that AT LEAST ONE item renders, not a specific count.
+- Do NOT guess CSS classes or inline styles in assertions.
+- Prefer querying by text content or role over testId — but if the source uses `data-testid`, prefer that.
+
 **Anti-Patterns (Do NOT Do These)**
 - Do NOT write tests that only call `render()` without any interactions or rerenders — these produce only mount-phase markers, which are USELESS for optimization measurement.
 - Do NOT create wrapper components with manual render count variables.