Get token-efficient answers from the Maze MCP

A guide to prompting your AI app so it pulls only the research data it needs.

Overview

When you ask your AI app a question through the Maze MCP, it fetches research data from your Maze workspace — study structure, results, participant answers, and transcripts — and reads all of it to write an answer. The more data it pulls in, the more tokens the answer uses.

Tokens are the units AI apps like Claude, ChatGPT, and Cursor use to measure how much text they read and write. Pulling in more than you need has three costs:

• It's slower. Bigger responses take longer to fetch and read.

• It burns through usage limits faster. Every app has a cap on how much you can do in a session or a billing period.

• It can crowd out the answer. If a single study fills the whole conversation, your app has less room left to reason, and answers get vaguer or cut off.

The good news: you control almost all of this through how you phrase your prompts. This guide covers the habits that keep answers fast, focused, and cheap. None of them require any settings changes — just clearer asks.

The one principle behind every tip

Ask for the narrowest thing that answers your question.

Almost everything below is a version of this. Point your app at one study instead of searching everything. Ask for the summary before the deep dive. Request the numbers when you need numbers, and the quotes only when you need quotes. Narrow to the people, dates, or questions you actually care about. When in doubt, start small — you can always ask for more.

General recommendations

1. Point straight at a study when you already know which one

If you paste a study's Maze link (or name it exactly), your app can go directly to it. If you describe it vaguely, your app has to search across your workspace first, then fetch — two steps instead of one.

• Do this: "Summarize the results of [paste the study link]."

• Instead of: "Find that checkout study we ran a while back and tell me how it went."

Keep the open-ended, searching questions for when you genuinely don't know what exists yet (see the next tip) — they're worth it then.

2. Use broad "what do we have?" questions on purpose, not by accident

Searching across all your research is one of the most useful things the MCP does — "What have we learned about onboarding?" pulls from many studies at once. Just know that a wide search touches a lot of data, so use it when discovery is the goal, and switch to a specific study once you've found the one you want.

• Great for discovery: "What research exists on [topic] in my [workspace] workspace?"

• Then narrow: "Open [that study] and walk me through the drop-off points."

3. Ask for the summary before the deep dive

Every study can be read at a summary level — the structure, the key stats, the completion and drop-off rates — without pulling in every participant's individual answers. For most questions, the summary is the whole answer.

• Do this: "Give me the high-level results and completion rate for [study]."

• Go deeper only when needed: "Now show me the individual responses to question 3."

Starting with the summary also lets you see how big a study is before you decide to pull everything, so you're never surprised by a huge response.

4. Ask for quotes only when you actually need the words

Transcripts and verbatim quotes are usually the single largest part of a research response. If your question can be answered with numbers — how many people did X, what percentage chose Y, where people dropped off, how long a task took — you don't need the transcripts at all.

• When you want the numbers: "What share of participants completed the task, and where did they struggle? Just the stats, I don't need quotes."

• When you genuinely want the words: "Pull three verbatim quotes where participants described the checkout as confusing."

Being explicit about "no quotes needed" or "just quotes on this one question" is one of the biggest savings you can make.

5. Narrow to the questions or steps you care about

You rarely need every question in a study. If you're only interested in one or two, say so, and your app can skip the rest.

• Do this: "Just the results for the pricing question and the final satisfaction score."

• Instead of: "Give me everything from [study]" — when you only wanted two answers.

6. Filter to the people and timeframe that matter

If you only care about mobile users, or the last 30 days, or people who completed the study, ask for that slice directly instead of having your app read the whole dataset and sort it out afterward.

• Do this: "Among mobile participants in the last month, what was the completion rate?"

• Instead of: "Show me all sessions" and then filtering in your head.

Time-based asks are especially worth scoping: "in the last 30 days" is far cheaper than "across all time," and usually more relevant.

7. Bundle related questions into one ask

If you have four things to ask about the same study, ask them together. Each separate message can make your app re-fetch the same study from scratch. One well-packed prompt reads the study once.

• Do this: "For [study]: give me the completion rate, the top drop-off point, the overall sentiment, and one representative quote."

• Instead of: four separate messages, each re-opening the same study.

8. Reuse what's already on screen

If your app just pulled a study into the conversation, follow-up questions can usually be answered from what's already there — no new fetch needed. You don't have to re-ask for the data; just ask the next question.

• Do this: "Based on what you just pulled, which segment struggled most?"

• Instead of: "Go get [study] again and tell me which segment struggled most."

Tips by type of study

Different research types carry their weight in different places. A quick orientation:

Unmoderated studies (surveys, prototype and usability tests)

Most of what you'll want lives in the stats, not the raw sessions. Completion rates, choice breakdowns, path success, average time on task, and drop-off points are all available at the summary level without reading a single individual response. Reach for individual sessions or transcripts only when you need to understand why behind a number — and when you do, narrow to the specific question and the specific slice of people.

Moderated interviews and AI-moderated conversations

Here the value is in the words, so transcripts are the main content — and they're heavy. Two habits keep this manageable:

• Start with the roster and summaries. Ask "How many interviews are there and what's the gist of each?" before pulling full transcripts. This tells you which conversations are worth reading in full.

• Read the heavy ones one at a time. If you want to go deep on a single interview, ask for just that one rather than every transcript in the study at once.

• Do this: "Summarize each interview in [study] in a sentence, then pull the full transcript of just the two most relevant ones."

• Instead of: "Give me every transcript in [study]" when you only needed two.

Quick reference

Putting it together

Here's the difference in practice.

Higher cost — vague and everything at once:

"Find our recent onboarding research and give me a full breakdown of everything participants said and did."

This makes your app search your whole workspace, pick a study, and pull every session and transcript — most of which you won't use.

Lower cost — same insight, scoped:

"In [paste onboarding study link], what was the completion rate and the top two drop-off points? Just the stats for now — no transcripts."

(then, once you've seen the numbers)

"Now pull two verbatim quotes from people who dropped off at the sign-up step."

Same understanding, a fraction of the tokens — and a faster, sharper answer both times.

A quick mental model

Before you hit send, a five-second check:

1. Do I know which study? → Paste the link.

2. Do I need numbers or words? → Ask for one, not both.

3. Do I need everyone, or a slice? → Name the slice.

4. Am I asking one thing or several? → Pack them into one prompt.

That's it. Scope the ask, and the speed and cost take care of themselves.

Need help connecting the Maze MCP, or running into unexpected results? See "What is the Maze MCP?" and the setup guides for your AI app.