> ## Documentation Index > Fetch the complete documentation index at: https://docs.deductive.ai/llms.txt > Use this file to discover all available pages before exploring further. # Asking better questions > The prompt patterns that consistently produce sharp investigations, plus a role-based prompt library you can steal from. Deductive is good at clarifying when it needs to, so you don't need to write the perfect prompt, but a few patterns reliably produce tighter, faster, more useful investigations. This page is the cheat sheet. ## The four patterns that always work ### 1. Anchor in time Vague: *"Why is the API slow?"* Anchored: *"Why did the API's p99 latency jump at 14:30 today and stay elevated?"* Time anchors give the agent a window to focus its evidence-gathering. "Yesterday between 11pm and 2am" beats "last night". "From the last successful deploy onward" beats "recently". ### 2. Anchor in change If you suspect something changed (a deploy, a config flip, a feature flag), say so. The agent will pull the change set and correlate against it as a first-class hypothesis. > "Did anything in PR #1234 cause the payment errors after we merged it?" > "We flipped the `cache_v2` flag to true at 12:00 PT. Did anything regress in the next hour?" > "Has any deploy in the last week touched code paths that the `auth-service` calls?" ### 3. Compare Comparison prompts force the agent to fetch baselines, which is almost always the correct framing for "is this normal?". > "Compare today's checkout error rate to the same time last week." > "Which services are noisier in their logs this week than last week?" > "How does the latency profile of `payments` look against the rolling 30-day baseline?" ### 4. Ask for the timeline, not the answer If you don't yet know what you're looking at, ask the agent to assemble the timeline first. The right answer often falls out of the timeline. > "Walk me through what happened to the payment service between 14:00 and 15:00, in order." > "Build me a timeline of every event. Deploys, alerts, infra changes. Touching the auth path yesterday." ## A prompt library by role Steal these. They work because they hit one of the four patterns above. > "What was the worst-looking thing in production last night between 11pm and 2am? Build the timeline first, then propose root cause." > "Has anything weird happened with the `` in the past 24 hours? Anchor on the last successful deploy." > "Summarize this week's incidents, group them by likely cause, and flag any that look like they share a root cause." > "Walk me through what changed between when `` fired and when it cleared." > "Three pages in two days for the same service. Are they actually the same incident or three different ones?" > "I merged PR #1234 yesterday. Did it break anything? Compare the metrics before and after the deploy." > "Has anyone else's recent PR touched the same files as #1234? If so, what did they change?" > "What does a normal deploy of the `` service look like in metrics? Compare against today's deploy." > "If I revert PR #1234, what else might I be reverting that depends on it?" > "Before I open this PR for review. Does it touch any code path that's recently been flagged in an incident?" > "Which services have the noisiest error logs in the last week, and is the noise growing?" > "Where are we paying the most for compute that's idle most of the time?" > "Show me services whose p99 has drifted up over the last 30 days without a corresponding traffic increase." > "Which alerts in `#prod-alerts` have fired more than three times this month? Cluster them by likely root cause." > "Compare the deploy frequency and revert rate of every service. Which teams are healthiest? Which need attention?" > "A customer says checkout has been slow for them since Tuesday. What does the data show for that customer specifically?" > "Which incidents in the last 30 days had visible customer impact, and how long did each one last?" > "What's the status of the open bugs filed against the payments team this quarter?" > "Build me a list of every recurring issue customers have raised about ``, grouped by likely cause." ## When the agent asks you a follow-up Sometimes Deductive will respond with a clarifying question instead of an answer. That's a feature, not a bug. It means the prompt was ambiguous in a way that would have led to a wasted investigation. Answer the question and the investigation continues. If you find yourself repeatedly answering the same clarifying question across different investigations, that's a signal that the underlying decision tree is worth reinforcing so the agent doesn't have to ask next time. See [Teach Deductive](/users/teach).