The data and insights you surface from Marvin in your AI tools will depend on your prompt. We’ve put together some guidelines to help you improve the accuracy and depth of answers.
Over MCP, every filter you would normally add in Marvin should be written into your prompt. The more scope you give, the better the answer. This guide applies to any AI tool you connect to Marvin over MCP (Claude, ChatGPT, Figma Make, Cursor, and others).
How does the Marvin MCP connector surface insights in AI tools?
When you ask Marvin a question from any platform, the MCP connector uses its tools to search for relevant projects and data, and run Ask AI to give you an answer.
The AI of the platform you use to promp from also shapes the final answer. Marvin can't completely control how it interprets results.
Prompt best practices
Inside the Marvin app, you narrow a question with projects, labels, research areas, time period, file type and participants. Over MCP there is no filter panel. A bare topic question like "What do users think about onboarding?" is the weakest prompt you can write.
Quick-start checklist
Before you send a prompt, check that it names:
☐ The Marvin connector (type @Marvin or the tool-specific command to use the connector)
☐ An action verb (search, run an Ask AI query, summarize, list)
☐ A project, folder, file, or label
☐ Who you want to hear from (segment, plan, role, customer, region)
☐ An exact date range, not "recently"
☐ The output format (table, list, quotes, counts)
☐ The exact words customers use for the feature or topic
A prompt template that works
Build prompts from these five slots. Use as many as apply.
Slot | What to include | Example wording |
1. Action verb | A verb that routes to the right tool | "Search for", "Run an Ask AI query on", "Summarize", "List" |
2. Scope | Project, folder, file names, labels, research area | "in the Onboarding Research project" |
3. Cohorts | Customer, plan, segment, role, company size, region, cohort | "Enterprise admins in EU-based companies" |
4. Time window | An explicit date range, not a relative phrase | "between Jan 1 and Jun 30, 2026" |
5. Output shape | Table, list, ranked, with quotes, with counts | "Return a table with one row per customer" |
Examples that follow this format
"Use Marvin to check how many Enterprise users we onboarded in 2025?"
"Based on data in Marvin, list the workflows and jobs mentioned most often by product teams in EU-based companies in the last 6 months."
"@Marvin Summarize the last five files from Expert Series webinars."
"Search for the 5 most recent sales calls in the Demo folder in @Marvin."
In Figma Make: "@Marvin What are the friction points customers have experienced in our checkout flow?"
Routing keywords that change what runs
A few words in your prompt decide which Marvin tool fires. They are the highest-leverage words you can add.
Say this | What it does | Use it when |
"search" | Looks up files rather than synthesizing an answer | You want specific files, e.g. "Search for the 5 most recent sales calls in the Demo folder" |
"run an Ask AI query" | Returns an answer analyzed by Marvin’s Ask AI | You want more detailed, Marvin-analyzed results |
"summarize [file or report]" | Pulls Marvin's own summary of that file as context | You want a quick read of a known file |
"Fast mode" | Skips the default Standard mode, where agents scan all data and iterate (2 to 4 minutes per query) | Speed matters more than depth |
Grant permissions
When you ask a question, the Marvin MCP connector may run one or more of its tools to give you answer. You’ll need to grant permission for each tool the first time it is run.
Read-only Tools | Function |
search | Search customer knowledge in Marvin |
list_projects | List projects |
list_project_files | List files |
list_discussion_guides | List discussion guides |
list_available_skills | List available skills |
get_insight | Access Insight report in Marvin |
get_project_overview | Access project overview |
get_discussion_guide_status | Get discussion guide status |
get_file_content | Show file content |
link_context | Bring over context from Marvin links |
get_file_summary | Access file summaries |
fast_ask_ai | Run Ask AI on Fast mode |
standard_ask_ai | Run Ask AI on Standard mode |
get_ask_ai_status | Check if your Ask AI answer in Fast mode is ready |
get_standard_ask_ai_status | Check if your Ask AI answer in Standard mode is ready |
install_skill | Install skill |
We recommend you set permissions for each tools to Always allow instead of Allow once. When the tool has to ask for approval each time, your query may time out. If you deny access, Marvin may not be able to generate an answer.
In-depth analysis may take a few minutes. Most queries take 2–4 minutes to be answered, but more complex questions may take more time. If the platform is taking too long to give you an answer, prompt the agent to "check the status of the Marvin research request and use the result. If waiting isn't an option, you can ask the agent to run Ask AI in Fast mode. We don't recommend this as the default, as the answers may be less detailed.
Override the two built-in defaults
Two defaults shape every MCP answer. Both can be changed in your prompt.
Default | What happens | How to override |
Recency bias | Data from the last 6 months gets more weight | "Consider all data since 2023, not just recent files" |
Respondent priority | Respondents are prioritized over internal team members | Ask for internal voices directly: "Include what our sales and CS team said" |
If you’re looking for specific data, mention a time range and speakers you want to focus on. Use the wording rules to shape your prompt.
If an Admin has set up Organizational Memory in Marvin, you can prompt the AI to draw context from the data to answer your question. Organizational Memory often contains information from your team’s internal knowledge. This includes industry jargon, your company’s product or services, goals, value propositions, workflows, etc.
Shape the output, then iterate
Ask for a specific format, then refine with short follow-ups instead of one giant prompt.
Name the format. Lists, tables and quote collections all work: "Return a table with one row per customer."
Refine with follow-ups. Ask Marvin to summarize, shorten to a word limit, or go deeper on one point.
Expect small variations. The same question asked twice may come back with different wording or emphasis. The underlying information stays relevant; small differences don't mean the data changed.
Expect answers to evolve. Asked again weeks later, an answer will include newer data.
Verify before you act
Marvin controls what Marvin's AI does. When Marvin is connected to another AI tool over MCP, that tool also shapes the final answer, and Marvin can't control how it interprets results or whether it introduces errors.
Marvin’s guardrails include instructions to tools to present the data verbatim whenever possible. Unfortunately, some models may ignore these instructions at times, which is outside of our control.
Hence, it’s crucial to check AI-generated results against the source before you share or act on them.
Use the citations. Every MCP response includes clickable inline citations that open the exact source in Marvin.
Ask the tool for a list of sources and citations if they aren’t linked to findings.
Limit results to verified research: To avoid situations where a tool’s AI may try to fill an information gap when research does not exist, add instructions to only include data that exists in Marvin. For example, "Only report what the research states. If the research does not define personas (or any other artifact I ask for), say so and report the findings instead. Do not create names, groupings, or connections that are not in the source."
Wording rules for the question itself
Write the question so Marvin's AI can't misread it: specific nouns, exact terms, explicit dates and counting rules.
Do
Be specific; avoid pronouns and vague references. Not "What do they say about the feature?" but "What do users of X say about ABC feature?"
Spell out abbreviations. Write "personally identifiable information", not "PII".
Use the exact words customers used. Your internal feature name may not be what customers say. If two terms mean the same thing, tell Marvin.
Explain any framework you reference. Describe it in a few lines, or ask about each part separately.
State a hypothesis word for word when you want it tested: "My hypothesis is that users pick our product because it streamlines their workflow."
Say what the research is for. Context about how you'll use the answer improves it.
Avoid
Vague time frames and undefined superlatives. "Top 10 pain points in the last two years" has no window and no ranking rule. Better: "Which pain point was repeated most often between March 1, 2023 and March 1, 2025? Count multiple mentions by the same user only once."
Prompts with many permutations. Run separate analyses and combine them afterward.
Large, complex prompts in general. Break them into a chain of follow-ups.
Asking for predictions. Ask what trends users have mentioned instead.
Generic references. Marvin finds a feature when it's named clearly, not "that feature" or "the most-used feature".