AI chat best practices

The AI chat reads every completed interview and answers in plain language. What you get back depends on how you ask. Use the practices below to get sharp, well-sourced answers, and to turn a good thread into a chart, a report, or a highlight reel.

Prompt for better answers

• Say what you want to learn, not just a keyword. "What made people hesitate at checkout?" gets you a synthesis; "checkout" makes the AI guess what you meant. A question gives it a target.
• Ask one thing at a time, then follow up. Each turn builds on the last, so start narrow and drill in ("now just the SMB customers," "show me the negative ones," "why did they say that?") instead of packing five asks into one prompt.
• Name the segment when you have one. The chat reads participant tags and screener answers. For guides with variants, it reads which variant each participant was assigned. For graded studies, it reads who passed or failed each task. Use those in the prompt: "How did first-time users differ from power users?", "How did people who saw the annual pricing variant react compared with the monthly one?", or "Why did people fail the sign-up task?"
• Check who reached the variant. Strella assigns a participant a variant when their interview starts, so someone who left early may never have seen it. The chat confirms from the interview itself. Ask "did everyone on that variant actually get to it?" when the difference matters.
• Anchor to your research objectives. Asking in the language of what you set out to learn ("Which of our objectives did the interviews actually answer?") keeps the analysis on the questions that matter to the study.
• Ask for the evidence. Every answer cites quotes inline, and you can click one to jump to that moment in the source interview. If a claim feels thin, ask "which interviews is that from?" or "show me the quotes."
• Push back when something looks off. The AI responds to challenges: "that seems high, double-check," "are there counterexamples?," or "you missed the Android users." Treat it like an analyst you can question.
• Set the scope with the study filter. Narrow to one study for depth. Widen across studies to spot patterns that span your research. In very large organizations a broad query may sample interviews rather than read every one, so narrow the filter when you need thorough coverage.

Make a chart

• Ask for the cut you want. "Chart how participants rated the onboarding flow" or "break down mentions of price by customer size." You can also click the Create a chart suggestion card and refine from there.
• Charts work best on countable data such as ratings, categories, coded open-ended answers, or task pass/fail. Open-ended sentiment is better explored as themes in analysis.
• Iterate on the view. Follow up with "make that a percentage," "split it by segment," or "only completed interviews" to reshape the chart.

Generate a report

When a thread is worth writing up, ask "build a report from this." Scoping it tightly ("a report on pricing feedback") produces a focused writeup instead of a sprawling one. The chat creates the report and opens it, and you refine the wording by asking the chat for changes. Full details: Build & share a report.

Create a highlight reel

Ask "make a highlight reel of the strongest moments about X." The chat compiles the clips and opens the reel. Save it to pin it, so it stops changing as new interviews come in. Keep reels short (four to six clips) and mix in a dissenting reaction so they read as honest. Full details: Create & share a highlight reel.

Say how long it runs, or how many clips it holds, in the same sentence: "a two-minute reel of pricing objections," "keep it under 90 seconds," "five clips on the checkout flow," "no more than three moments."

The chat treats a number of clips as exact. It aims at a length and may come out slightly over or under it. It picks whole moments and never cuts into the middle of an answer, so a reel asked for two minutes comes back near two minutes. Trim individual clips in the reel editor when you need an exact runtime. Twenty-five clips is the most any reel holds.

The chat tells you when it cannot hit what you asked for. It picks only from moments already found in your conversation, so asking for five clips when three of them match gets you three clips and the reason why. The same goes for a length no combination of moments can reach. The chat never claims a reel met a length or a count it missed.

Example prompts

Things to know

• Only completed, non-excluded interviews are in scope. The chat leaves drafts, in-flight interviews, and interviews excluded from analysis out of every answer, chart, report, and reel.
• Answers cite their sources. Click any quote to jump to that moment in the source interview's transcript.
• Reports and reels aren't edited inline. Ask the chat to revise a report, and edit a reel's clips from the reel itself.

Related