Minds Team

MCP Workflow Examples for AI Market Research

[zh] Block-first MCP workflows for cohesive multi-question Studies.

These examples assume you have connected the Minds MCP server for ChatGPT, Claude, and Cursor. If not, follow the Minds MCP setup guide first.

Guided Study Confirmation Workflow

For a broader task, multiple questions, a visual asset, structured outputs, or an explicit research method, the assistant follows this sequence:

User request
  -> create_study_plan (draft only)
  -> present captured intent, main source, questions, methods, and outputs
  -> ask every returned confirmation question
  -> revise_study_plan if the user answers or changes anything
  -> explicit user confirmation of the exact revision
  -> start_study_run
  -> get_study_status with the returned runId until collection settles
  -> read_study_summary

模型不能默默批准自己提出的问题或方法。简单 Studies 应保持简单。请检查 list_research_methods：Conjoint 和其他可用定量方法会执行已注册的设计与计算；高级方法需明确同意。确认计划前先审阅替代方法。对已关联问题的素材，get_study_heatmap / start_study_heatmap 可读取或启动独立视觉分析。仅报告实际返回的证据。

The original create_study_plan.request is planner input and is not sent verbatim to Minds. The preview labels the exact proposed respondent-visible question text; after confirmation, get_study_run exposes the confirmed or server-prepared question set as a visibility audit. By contrast, treat all text in a direct ask_study.question as respondent-visible input: the system may classify or reformat it, but any part can reach the Minds and influence their answers. Keep planner and MCP-client orchestration notes outside direct questions.

At the execution boundary, a plan_limited response means nothing was started: explain the Study-answer limit and required upgrade. During polling, status: plan_limited means the processor preserved partial answers but stopped the remaining questions at the paywall. Never call either case complete, and never hide it behind a generic error. After an upgrade, use a follow-up run for only the unanswered questions.

已确认的运行会像一位受访者填写问卷那样作答：题目按顺序执行，每个 Mind 只能看到自己在本次运行中先前的回答，因此会像真实调查一样出现顺序效应。当研究设计需要控制顺序效应时，请在不同的运行之间轮换题目顺序。如需跳题逻辑，请为题目设置 askIf，指定前面的一道选择题及其选项；未被提问的 Minds 不计入该题结果，没有任何人被提问的题目会报告 not_asked，而不是失败。运行结束后，answerConsistency 会标记与同一 Mind 其他回答相矛盾的回答。这些标记仅供审阅，回答本身永远不会被修改。

This lifecycle plans research, not every Study action. A standalone “export this” request routes to export_study. “Show the same result differently” reads the existing Study or summary and changes presentation without queueing new respondents. Only create a new plan when the requested output changes what evidence the Study must collect.

Consumer Research Workflow

A typical workflow for a brand team researching a product launch:

Step 1: Create Consumer Personas

"Create five consumer personas for our skincare launch:
 - 'Lena, 22, Berlin' — Gen Z student, eco-conscious, discovers brands on TikTok
 - 'Maya, 19, London' — retail worker, trend-driven, shops based on peer recommendations
 - 'Aisha, 24, NYC' — junior designer, budget-conscious, cross-references Reddit reviews
 - 'Sarah, 35, Munich' — part-time teacher, mom of two, safety-first buyer
 - 'Jessica, 38, Chicago' — marketing manager, mom, trusts dermatologists over influencers"

Step 2: Organize into Demographic Audiences

"Create an Audience called 'Gen Z Women (18-25)' with Lena, Maya, and Aisha.
Create another Audience called 'Millennial Moms (30-42)' with Sarah and Jessica."

Step 3: Create a Research Study

"Create a Study called 'Sustainable Skincare Launch Study' with both Audiences"

Step 4: Run Survey Questions

"Plan one launch-research question block for the Study with these sections:
1. Price sensitivity: On a scale of 1-10, how likely are you to switch to a sustainable skincare brand if it costs 20% more?
2. Discovery: Where do you typically discover new skincare products?
3. Switching barriers: What would make you stop buying from your current skincare brand?
Show me the complete plan for confirmation, then run all three questions together."

Step 5: Analyze & Export

"Show me the analytics for the Sustainable Skincare Study"

"Export the Study as a PDF report for the brand team"

验证受众

要检查受众的回答与已发布的真实调查有多接近，助手会按以下顺序操作：

创建或选择受众（至少 10 个已就绪的 Minds）
  -> validate_audience（查找最匹配的已发布调查，或使用你指定的 benchmarkIds）
  -> 使用 batchId 调用 get_audience_validation，直到 status 为 completed、failed 或 cancelled
  -> 读取综合分数、每项调查的分数和 95% 区间，以及被排除的问题

对于直接调用工具的 MCP 客户端，validate_audience 和 get_audience_validation 可以按名称调用。它们尚未出现在助手的工具发现列表中；同样的功能现在可以通过 v1 API 和受众的验证标签页使用。

一次运行通常需要 10 到 60 分钟，因此助手会定期轮询状态，而不是在一次调用中等待。它会报告每项调查的分数、区间以及调查对象，并向用户展示被排除的问题及其原因。不带 batchId 调用 get_audience_validation 时，会返回受众的整体效度和本月剩余的套餐内验证次数。

开始之前，助手会确认受众已稳定：readiness.fullyTrained 为 true，且研究阶段不再是 researching。更早启动的验证会按受众当时的状态评分。运行期间，Minds 根据各自的画像作答，不会检索其知识文档，也不会搜索网络。报告分数时，助手会说明该分数是 100 减去每个答案选项的平均绝对差距，而不是答对的比例，因此高分也可能掩盖某个选项上的较大偏差。每个问题的人类与合成份额以及噪声下限可通过 v1 API 获取。

说明部分人群的匹配结果之前，助手必须先读取 respondentScope。例如，护士调查可能从包含16个医疗工作者 Mind 的 Audience 中选出12个。展示百分比时应同时注明“16个 Mind 中的12个，护士”，不要将其描述为所有医疗工作者的验证结果。结果提示会区分由群组数据确认的资格、由更广泛个人档案支持的资格、不确定情况以及明确不匹配的情况。

资格判断在收集回答之前完成，并独立于回答内容。个人证据可以补充群组信息的缺口，但汇总分布不能证明某个人符合条件。只有已确认或有证据支持的成员参与，且至少需要10个。未达到门槛时，应解释为什么没有进行调查。模型负责选择对象，匹配率则通过数值比较计算。费用按实际选中的回答者计算，综合分数不包含子集结果。即使部分 Mind 也属于其他 Audience，解释仍应限定于本次 Audience。

示例提示：

"用已发布的调查验证我的 'German SaaS buyers' 受众，并告诉我它有多接近。"

Use Cases

Product Concept Testing

Create Studies with target consumer Audiences to test product concepts, packaging, and naming before investing in production. Compare reactions across Audiences instantly.

Pricing Research

Survey synthetic consumers on willingness to pay at different price points. Identify the sweet spot where value perception meets margin targets.

Brand Perception Audit

Build Studies containing your target Audiences. Ask about brand awareness, trust, and purchase intent in one cohesive question block. Compare your brand against competitors through consumer eyes.

Campaign Message Testing

Test ad copy, taglines, and visual concepts with synthetic target audiences before committing media spend. Identify which messages resonate with which segments.

Market Entry Research

Entering a new market? Create consumer personas for the target region and test product-market fit, cultural sensitivities, and channel preferences.

Example Conversations

Quick Consumer Insight

You: "I need to understand how Gen Z and millennials in Europe feel about subscription-based skincare. Create one Audience for each segment and plan one cohesive Study asking what would convince them to subscribe."

AI Assistant: Drafts the Minds and audience structure, creates a Study, presents the exact proposed research question(s), and requests confirmation when the request expands into a broader Study. After execution, it reports the actual returned grouped responses without pre-writing a conclusion.

Expert Consultation

You: "Talk to my Brand Strategy expert about positioning a premium organic baby care line in the German market."

AI Assistant: Uses chat_with_mind to query the expert Mind and presents its actual response, preserving any returned citations. It labels the output as synthetic expert perspective rather than independent legal or regulatory advice.

Competitive Analysis

You: "Create a Study with 'Loyal Customers' Audiences for our top 3 competitors. Plan one block asking what they love most about their current brand and what frustrates them."

AI Assistant: Creates private synthetic Audiences, confirms the comparative multi-question research plan, runs the Study once, and reports only differences supported by the returned responses and calculations.

Automation guardrails

  • Keep new Audiences and Studies private unless the user explicitly requests a public link.
  • Planning does not start research. Confirm the exact latest revision before start_study_run.
  • Poll status tools instead of treating a timeout or elapsed duration as completion.
  • Require explicit confirmation immediately before lifecycle deletion actions.
  • Preserve citations, resource IDs, workspace/shared links, and download links exactly.
  • Treat plan_limited before execution as “nothing started” and status: plan_limited during execution as partial, preserved work.
  • Use list_research_methods before promising a named method; only executable: true is runnable.
  • Do not infer customer findings in advance. The example prompts in this guide describe workflows, not guaranteed conclusions.

For a full routing and safety contract, see the MCP operating guide for agents.

如果缺少某个问题或 Mind 的回答，研究运行可能以 status: "partial" 结束收集。此状态为终止状态，请停止轮询并说明缺失情况，不要将结果称为完整。progress.pct 衡量已结束的问题数，而非受访者回答覆盖率。新问题输出在 outputData.responseCoverage 中保留 expected、received 以及缺失的 Audience/Mind 成员关系。progress.partial 统计已记录覆盖不完整的问题数。旧结果缺少覆盖数据并不证明所有成员都已回答。保留有效的原始回答及其结果，不作修改；另行重试前先检查缺失内容。

等待就绪后继续

在网页版 ChatGPT Work 或桌面版 Work 的 Cloud 模式中，请 Minds 在受众就绪或已确认的研究结束时通知你，也可在受众就绪后继续已批准的研究。请先审核并确认受众草稿和研究计划：最初的请求不能批准尚未查看的计划，事件也不会授权额外问题、新运行或额外支出。通知渠道由主机决定，不保证发送短信。

受众事件会等待新建 Minds 可用；等待审核的预览不算就绪。研究事件会等待结果综合已保存。回答不完整、临时摘要、失败、取消和套餐限制均报告 ready=false。用 get_study_status({ study: { id: studyId }, runId }) 读取原运行并保留其标识；不要因收到事件而重新提交。

订阅最长 24 小时，客户端必须在 refreshBefore 前续订。签名密钥和回调网址加密保存。每次交付会重新检查连接、账户规则和资源访问权。待处理状态每分钟检查一次；回调验证及重试可能增加延迟。不提供历史重放。如果订阅失败，请保留工作区链接并查询状态；订阅成功前不要承诺通知。

监控任务应以研究名称和用户要求的阶段命名，不要在结果未知时承诺成功。通知应使用用户的语言，开头先说明结果和研究名称，紧接着给出工作区链接及下一步操作。ChatGPT 的邮件预览可能截断完整消息，因此开头本身就应清晰有用。仅在已保存结果能够验证时，补充简短发现以及受访者数或问题数。除非用户要求，普通完成通知不要包含事件名、内部标识符或 JSON 字段。

对于等待确认的草稿，应说明“可供审阅”且研究尚未开始；在完整审阅对话中保留准确的草稿 ID、修订版本和确认问题。遇到失败或部分完成时，读取原始操作的 get_audience_creation_progress({ operationId }) 或 get_study_status({ study: { id: studyId }, runId })。说明返回的原因、已保存的进度、缺失内容及可用的下一步操作；如果没有原因信息，应如实说明。不要编造原因或自动重复研究。只通知用户要求的阶段；已获授权的中间步骤应继续执行，除非需要审阅或受到阻碍，否则不额外通知。邮件模板和通知渠道由宿主控制，因此不能承诺邮件内的直接按钮或短信。

audience.creation.finished → operationId · study.run.finished → studyId, runId