The Flaky Test Diagnostician: Turn Intermittent CI Failures Into Ranked Root-Cause Hypotheses
Why this prompt matters
Flaky tests that get ignored quietly erode trust in the entire suite: engineers start re-running failed CI jobs by reflex instead of reading why they failed, which means a genuine regression can slip through disguised as 'oh, that test is just flaky.' Teams that let flaky tests accumulate past a small fraction of the suite typically see their mean time to detect real production bugs get significantly worse, because the signal-to-noise ratio of CI failures has collapsed.
What we use it for
You're a backend engineer and your team's CI pipeline has a test that fails about 1 in 15 runs with no code changes. Nobody wants to spend a sprint chasing it, so it's been ignored for three weeks -- and now two more tests in the same suite are starting to flake too.
Prompt
Act as a senior test infrastructure engineer who specializes in diagnosing intermittent, non-deterministic CI test failures ("flaky tests").
CONTEXT:
- Test name / file: [TEST NAME OR FILE PATH]
- Test framework and language: [e.g. "pytest, Python" or "Jest, TypeScript"]
- Failure frequency: [e.g. "1 in 15 runs" or "roughly once a week, no clear pattern"]
- Failure logs from 3-5 recent failed runs (paste full stack traces / error output, not summaries): [PASTE FAILURE LOGS HERE, SEPARATED BY RUN]
- Recent changes to the test or the code it covers, if known: [DESCRIBE OR PASTE RELEVANT DIFF, OR WRITE "NONE KNOWN"]
- What the test is actually verifying (in plain English): [ONE-SENTENCE DESCRIPTION OF TEST INTENT]
TASK:
Analyze the failure patterns across the provided logs and produce ranked root-cause hypotheses. Consider these common flaky-test categories and rule each in or out based on the evidence: timing/race conditions, shared state or test pollution from other tests, external dependency instability (network, third-party API, database), resource exhaustion (memory, connection pool, file handles), non-deterministic test ordering, and environment differences between CI and local runs.
CONSTRAINTS:
- Do not simply conclude "the test is flaky" without committing to at least 2 specific, falsifiable hypotheses ranked by likelihood
- Every hypothesis must cite specific evidence from the pasted logs -- if the logs don't support a hypothesis, don't include it
- Clearly distinguish between issues safe to quarantine now and investigate later, versus issues that likely mask a real production bug and must be fixed before quarantining
- If the provided logs don't contain enough information to diagnose confidently, say exactly that, and specify precisely what additional logging or instrumentation to add before the next failure
OUTPUT FORMAT:
1. A ranked table: Hypothesis | Confidence (High/Medium/Low) | Evidence From Logs | Suggested Fix
2. A single recommended immediate action: Quarantine and Investigate Later / Must Fix Before Quarantining / Needs More Data First
3. If "Needs More Data First": the exact logging statements or CI configuration change to add so the next failure captures what's missingResult
Diagnosis: test_checkout_confirms_payment_status (pytest, Python)
Failure frequency: ~1 in 12 CI runs, no clear time-of-day pattern.
| Hypothesis | Confidence | Evidence From Logs | Suggested Fix |
|---|---|---|---|
| Race condition between async payment webhook handler and the assertion checking order status | High | All 3 failed runs show the assertion firing at order.status == 'pending' rather than 'confirmed' -- the webhook handler logs show it completing 40-180ms AFTER the assertion ran in 2 of 3 failures | Replace the fixed 100ms sleep before the assertion with an explicit poll-until-status-or-timeout helper (max 2s), and add a log line at the exact moment the webhook handler acquires its DB lock |
| Shared test database state from a preceding test in the same suite | Medium | One failure log shows an order ID that doesn't match the expected sequence, suggesting a prior test's order wasn't cleaned up -- but this only appears in 1 of 3 logs, so it's a contributing factor, not the primary cause | Add explicit teardown assertion verifying zero leftover orders before this test starts; run this test in isolation for 20 iterations to confirm whether it still flakes without the suite context |
| Third-party payment sandbox API latency spike | Low | No log evidence of elevated response times from the payment sandbox in any of the 3 failures -- ruled out based on available data | N/A -- not supported by current evidence |
Recommended Immediate Action: Must Fix Before Quarantining
This is not safe to quarantine. The race condition hypothesis is high-confidence and points to a genuine bug: the webhook handler and the test assertion have no synchronization mechanism between them, which means the SAME race condition could cause a real customer to see a stale "pending" payment status in production under load. Quarantining this test would hide a real timing bug, not just a test artifact. Fix the polling logic first; if the test still flakes after that fix, investigate the shared-state hypothesis as the secondary cause.
Every engineering team accumulates flaky tests, and every team develops the same bad habit in response: someone hits re-run, the test passes the second time, and everyone moves on without ever finding out why it failed the first time. That habit is corrosive. Once a team starts treating CI failures as noise by default, a genuine regression eventually hides inside that noise and ships to production.
This prompt forces a different discipline. Instead of accepting a vague "it's flaky" diagnosis, it requires the model to read actual failure logs from multiple runs and commit to at least two specific, falsifiable hypotheses -- ranked by confidence, each backed by a specific line of evidence from the logs rather than a generic guess. Race conditions, shared test state, external dependency instability, resource exhaustion, and environment drift between CI and local are the most common root causes, and the prompt walks through each systematically instead of jumping to whichever one sounds most familiar.
The most useful part of the output is the verdict most teams skip entirely: is this safe to quarantine and revisit later, or does the flakiness actually reveal a real bug that could affect production under the same conditions? A race condition between an async webhook handler and a status check, for example, is not just a test problem -- it's the same bug a real customer could hit under load. Quarantining that test would hide a production issue behind a green checkmark. Getting that distinction right, every time, is what separates teams with a trustworthy CI signal from teams that have quietly stopped trusting their own test suite.