The Stress Test is not a grade — it's a peer review. It reads your prompt as a whole and tells you where the blocks pull against each other, where context is missing, and where AI is most likely to guess.
Anything client-facing, legal, financial, or "you only get one shot". Otherwise, the Lab score is enough.
Steps
Open your prompt in the Lab
From the Composer use "Open in Lab", or paste your prompt directly into the Lab input.
Click "Run Stress Test"
It takes ~10–20 seconds. While you wait, re-read your prompt cold.
Read tensions before suggestions
A "tension" is two parts of your prompt fighting each other. Resolve those first — the rest often dissolves.
Apply one fix, re-run
Don't batch every suggestion. Apply one, see what changes, then decide on the next.
Save the improved version
When the test comes back clean, save the prompt back to your Library so future-you starts here, not at the rough draft.
It doesn't run the prompt against a model. It doesn't grade you. It surfaces the things a careful colleague would flag if they read your prompt over your shoulder.