Research Insight Synthesis

Use case:

• Validating the preliminary synthesis
• Refining the logic and clarity of synthesis structure
• Identifying patterns from visual data
• Producing summaries or key points (e.g., design principles) based on uploaded data

Output:

• Logically structured synthesis documents

Survey

Input:

• Cleaned raw CSV data. (Important: clean the data before using it with LLMs to avoid inaccurate analysis.)

Steps:

• Create a project to isolate data and avoid hitting chat limits quickly
• Ask the LLM to describe the CSV to ensure correct interpretation
  • (E.g., prevent close-ended questions from being misread as open-ended during sentiment analysis)
• For slicing or subgroup analysis, create a new chat
• Carefully review results to detect hallucinations or misinterpretations

Output:

• Descriptive analysis (e.g., means, distributions)
• Inferential analysis (e.g., correlations between variables)
• Full-blown report
• Charts and graphs

🪜General Process

1. Create a new project in ChatGPT / Claude / NotebookLM
2. Upload data
3. Enter project instructions (e.g., "output in Bahasa Indonesia," or "include supporting evidence from uploaded documents for every insight")
4. Add synthesis notes or research context
5. Review outputs to avoid hallucinations or bias. Examine claims and verify against the uploaded data
6. Edit results to better align with organizational context and audience. Responsibility for final outputs remains with you.

Impact & Reflections

⇢ Impact

• Enabled fast understanding of complex or unfamiliar contexts
• Delivered insights in formats that are easy for other teams to absorb (text or audio)
• Provided quick snapshots of available data, which could be enriched by the team’s cultural context
• Helped generate new perspectives, including from unconventional data like visuals

⇢ Reflections

Do's:

• Carefully review input documents, especially key info or quotes that must appear in the synthesis
• Add contextual background (e.g., business process context or niche abbreviations) into input docs (for NotebookLM) or chats (for ChatGPT/Claude)
• Feed Miro data into LLMs as images (screenshots or PDFs), not as plain text, to preserve meaning-based layout
• Guard against hallucination by:
  • Using projects in ChatGPT, or using tools like NotebookLM/Claude that don't link chats
  • Asking the AI to cite supporting evidence for every claim (quotes or document pages)
• Iterate gradually. Ask for broad outputs first (e.g., "10 dominant behavior patterns"), then dive deeper into the promising ones

Don’ts:

• Request synthesis without thinking through the goal; the result will often be generic
• Upload uncleaned raw data directly into LLMs for synthesis