Running a Research Mode Query Step by Step
Research mode is only as good as the question you give it and the read you give its report afterward.
//
Search across all documentation pages
Research mode is only as good as the question you give it and the read you give its report afterward.
This page walks through a full run, start to finish - framing the question, starting the research, and reading what comes back - so the process feels concrete rather than abstract.
Research mode works in three phases from your side of the conversation: framing the question, letting the investigation run, and reviewing the report it returns.
Each phase has its own habits that separate a useful research report from a shallow one.
Framing matters most, because a vague question gives Research mode little to actually decompose and investigate.
Once started, the process runs on its own - your job shifts from directing it to waiting and, later, reading critically.
The review step is not optional: a cited report is only trustworthy once someone has actually checked a sample of those citations.
Quick-reference shape for a Research mode run.
1. Frame a specific question with real constraints (who/what/when/scope).
2. Start the research task from a normal Claude.ai conversation.
3. Let it run - it works through sub-questions and sources on its own.
4. Read the returned report, starting with the summary, then the citations.
5. Spot-check 2-3 citations that matter most to your decision.
6. Follow up with a narrower question if the first report was too broad.When to reach for this:
Here's a complete run from framing through review, using a realistic business question.
Step 1 - Frame the question:
Weak framing: "Tell me about project management tools."
Strong framing: "Compare three current project management tools -
Asana, Linear, and Monday.com - on pricing tiers, seat limits, and
whether each offers a free tier as of today. Cite each vendor's own
pricing page."
Step 2 - Start the research task:
Send the strong framing above as your message and start Research mode.
Step 3 - Let it run:
Research mode plans sub-questions (one per tool), searches for
current pricing information on each, and reads what it finds.
This step takes noticeably longer than a normal reply - that's expected.
Step 4 - Read the returned report:
"Report: Project Management Tool Pricing Comparison (as of [date])
- Asana: Free tier available (up to 10 users); paid tiers from
$X/user/month. [Source: asana.com/pricing]
- Linear: No unlimited free tier; paid tiers from $Y/user/month.
[Source: linear.app/pricing]
- Monday.com: Free tier available (up to 2 seats); paid tiers from
$Z/user/month. [Source: monday.com/pricing]
Note: pricing pages for all three list monthly and annual options;
this report cites monthly list pricing unless noted."
Step 5 - Spot-check citations:
Open at least the two vendor pricing pages most relevant to your
decision and confirm the figures still match what the report says.
Step 6 - Follow up if needed:
"Now break down Asana's paid tiers in more detail - what's included
at each level?"What this demonstrates:
| Weak framing | Why it's weak | Stronger version |
|---|---|---|
| "Tell me about X" | No comparison, no scope, no constraint | "Compare named options A, B, C on specific attributes" |
| "What's the latest on Y" | No time window, no source expectation | "What changed about Y in the last quarter, citing primary sources" |
| "Is Z a good idea" | Opinion question, nothing to research | "What are the documented trade-offs of Z, based on independent sources" |
A useful reading order is: read the summary first for the overall shape of the answer, then work through each cited claim that actually matters to your decision, then check whether the report flagged any disagreement between sources.
Claims that don't matter to your decision don't need the same scrutiny as the ones that do - reading a research report critically doesn't mean opening every single citation.
| Alternative | Use When | Don't Use When |
|---|---|---|
| Regular conversation | The question is well-known, opinion-based, or doesn't need comparing sources | You need a defensible, checkable answer for a real decision |
| Research mode with connectors | The question needs internal company knowledge alongside public sources | No connectors are set up yet for the sources you'd need |
| Manual research | You need to personally verify every step for a high-stakes decision | Time is limited and a first-pass synthesized report would still help |
Specific enough that Research mode can turn it into concrete sub-questions - naming entities, attributes, and a time window is usually enough.
Vague, open-ended questions tend to produce broad, shallow reports.
Research mode runs the investigation on its own once started rather than pausing for step-by-step direction; if the framing was off, the more effective move is usually to let it finish and then follow up with a narrower question.
Start with the summary for the overall shape of the answer, then read the cited detail behind any claim that actually matters to your decision.
No.
That's useful information itself - it tells you the topic may not have strong, independent public coverage, and you should weigh the report's confidence accordingly rather than treating it as settled.
Yes - a follow-up in the same conversation can drill into one part of the report without you needing to restate the original scope.
This usually traces back to framing - a question without named entities, specific attributes, or a time window gives Research mode more ground to cover and less to narrow in on.
Yes.
Planning sub-questions, searching multiple sources, reading them, and synthesizing a cited answer is genuinely more work than a single-pass reply, so a longer wait than usual is expected.
Not automatically.
Citation count matters less than citation diversity - several citations that trace back to the same original source don't add independent confirmation, even though the report looks well-sourced.
Rewrite the original question to be more specific - name the exact entities, attributes, and time window you care about - rather than assuming the feature simply underperformed.
Yes, and it's worth doing - a consistent question template for a recurring research task (like a monthly competitor check) makes reports easier to compare over time, and can be packaged into an Agent Skill if the team runs it often.
Only after a short review pass - checking citation quality, noting any dated claims, and confirming nothing in the report surprised you enough to double-check first.
Stack versions: Written against the Claude model lineup current as of ~June 2026 - Claude Fable 5, Claude Opus 4.8, Claude Sonnet 5 (the default), and Claude Haiku 4.5. Model names, pricing, and product features move quickly - verify current specifics at platform.claude.com/docs before relying on them.
Reviewed by Chris St. John·Last updated Jul 16, 2026