You have a browser tool connected and the app is running at {URL}.\n\nSweep ONLY these surfaces: {LIST THE 3 TO 5 PAGES OR FLOWS}.\n\nFor each surface:\n1. Load it and record the console errors and failed network requests you see.\n2. Exercise every interactive control once. Note anything that throws, hangs past 3 seconds, or leaves the UI in a state the user cannot exit.\n3. Resize to 375px wide and repeat step 2.\n\nReport each finding as: surface, exact steps to reproduce, what you expected, what happened, and the console or network evidence. Rank by whether a paying user would hit it.\n\nDo not read or modify source files. Do not fix anything. Stop after the listed surfaces.
Run: git diff {BASE_BRANCH}...HEAD --stat\n\nFrom that diff alone, list the user-facing behaviours that could plausibly have changed. Be specific about which file drove each one.\n\nThen, for each behaviour, use the browser tool to verify it still works at {URL}. Test the behaviour as a user would, not by reading the code.\n\nReturn a table: behaviour, file that changed it, pass or fail, and for failures the exact reproduction steps.\n\nIf the diff touches nothing user-facing, say so and stop rather than inventing checks.
A user reports: {PASTE THE REPORT VERBATIM}\n\nStep 1. Reproduce it at {URL}. Tell me the exact sequence that triggers it. If you cannot reproduce it in 10 attempts, say so and list what you tried and what differed from the report.\n\nStep 2. Minimise. Remove one step at a time and re-test until you have the shortest sequence that still triggers it.\n\nStep 3. Characterise. Does it depend on viewport, account state, data volume, or timing? Test each one.\n\nReturn the minimal reproduction, the conditions it needs, and the console or network evidence. Do not propose a fix yet.
Target the form at {URL}.\n\nFor every field, test and record the result for: empty, whitespace only, one character, the stated maximum length, one over the maximum, a leading and trailing space, a Unicode emoji, an HTML tag, and a SQL-style quote character.\n\nThen test the form as a whole: submit twice rapidly, submit with the network throttled, and submit after changing a field and pressing browser back.\n\nReport any case where the error message is missing, wrong, or exposes an internal detail, and any case where invalid data was accepted.\n\nRecord what happened. Do not change validation rules.
Walk the flow {NAME THE FLOW} at {URL} using only the keyboard.\n\nRecord: every element that cannot be reached by Tab, every place the focus ring is invisible, every point where focus jumps somewhere unexpected, and every modal that does not trap focus or does not return focus on close.\n\nThen check each interactive element for an accessible name, and each image for alt text that says something useful rather than repeating the file name.\n\nReport findings with the element selector and what a keyboard or screen reader user would experience. Order by how badly it blocks completion of the flow.
Run the test suite {COMMAND} five times in a row. Record which tests fail on at least one run but not all five.\n\nFor each flaky test, read it and classify the cause as one of: a fixed wait that is too short, a dependency on test execution order, shared state that is not reset, a real race in the code under test, or an external service.\n\nQuote the specific lines that led you to each classification.\n\nReturn a table ordered by how often each test failed. Recommend a fix for each, but do not change any file yet.