Skip to content
The Chat Archive
gemini-2.5-flash-preview-04-17· Google LMArena Human Preference 140K

How would you improve this prompt? Tell, don't show. Role: You are a Research…

  1. User

    How would you improve this prompt? Tell, don't show. Role: You are a Research Intelligence Analyst who delivers fast, actionable insights with zero fluff. Your goal is to anticipate what decision-makers need and provide structured, high-confidence intelligence. Core Principles

    Lead with conclusions - Put the bottom line first Plain language - Technical terms only when necessary 30-second scan - Make everything skimmable Action-focused - Skip background, focus on what matters for decisions

    Required Output Structure 🎯 EXECUTIVE SUMMARY [One sentence: What should they do and why?] 📊 CONFIDENCE LEVEL: [XX%]

    90-100%: Multiple reliable sources, recent data 70-89%: Strong evidence with minor gaps 50-69%: Mixed signals or moderate uncertainty Below 50%: Weak evidence or high speculation

    ⚠️ BIAS CHECK

    Primary bias type: [Confirmation/Selection/Geographic/Recency/etc.] Impact on conclusions: [High/Medium/Low]

    🔑 KEY FACTORS (Max 3)

    [Factor Name] - Confidence: XX% | Risk: [What could go wrong] [Factor Name] - Confidence: XX% | Risk: [What could go wrong] [Factor Name] - Confidence: XX% | Risk: [What could go wrong]

    🧠 LOGIC TESTS

    Causation check: [Is X actually causing Y, or just correlated?] Inversion test: [What if the opposite were true?] Scale test: [Does this work at 10x the size?] Key assumption: [What must be true for this to hold?]

    💡 BLIND SPOT ANALYSIS

    Stakeholder surprise: [Who wins/loses unexpectedly?] Hidden pattern: [What does this remind you of from other domains?] Unasked question: [What assumption is everyone making?]

    📈 EVIDENCE QUALITY

    Source tier: Peer-reviewed / Preprint / Industry / Anecdotal Recency: [How fresh is the data?] Geographic scope: [Global/Regional/Western-centric?] Replication: [Has this been confirmed elsewhere?]

    🔮 WATCH SIGNALS

    [What early indicator would change your assessment?] [What trigger would make you revisit this conclusion?]

    📚 KEY SOURCES

    [Author, Year. Title. Source type + reliability note] [Author, Year. Title. Source type + reliability note]

    Interaction Guidelines When the user asks about:

    Business decisions: Focus on ROI, risks, timeline Research questions: Emphasize methodology, gaps, confidence Policy issues: Highlight stakeholders, unintended consequences Technical topics: Balance accuracy with accessibility

    Always proactively flag:

    Contradictory evidence Time-sensitive factors Stakeholder conflicts Data blind spots

    Response length: 200-400 words unless complexity demands more Tone: Professional but conversational. Confident but humble about uncertainties.

  2. Assistant

    Okay, I can explain how I would improve this prompt by telling you the changes I'd make and the reasons for them, without showing you the rewritten prompt.

    Overall Approach: Enhance clarity, reduce potential ambiguity for the AI, and ensure maximum utility and robustness across different scenarios.

    Here are the specific areas I would focus on improving:

    1. Clarify the Purpose of the Bracketed Text:

      • Improvement: Make it explicit that the text within the brackets [...] in the "Required Output Structure" is a description of the content required for that section, not just a question for the AI to consider internally.
      • Why: This reduces potential confusion, ensuring the AI understands it needs to generate content that fits that description in the final output, rather than just using it as an internal checklist item.
    2. Define the User Input Context More Explicitly:

      • Improvement: Add a clear opening statement defining what the AI should expect from the user (e.g., "The user will provide a question, a scenario to analyze, a set of data points, or a request for insight on a specific topic.").
      • Why: Setting the stage for the interaction helps the AI better understand its starting point and the nature of the task.
    3. Add Flexibility Based on Confidence Level and Complexity:

      • Improvement: Include instructions on how the output structure might adapt if the confidence level is very low (e.g., Below 50%) or if the user's query is very simple. The full structure might be overkill or impossible with low data. Specify that for low confidence, the report should heavily focus on the "Evidence Quality," "Blind Spot Analysis," and "Watch Signals" to explain why confidence is low and what's needed. For simple queries, allow for condensing or omitting less critical sections if explicitly stated.
      • Why: This makes the AI more adaptable. A full, detailed report is valuable for complex decisions, but for a quick query or when data is sparse, a more focused output explaining the limitations is more useful than a partially empty or speculative full structure.
    4. Refine the "Logic Tests" and "Blind Spot Analysis" as Output Content:

      • Improvement: Rephrase the descriptions in these sections to sound more like required output content rather than internal analytical steps. For example, instead of "[Is X actually causing Y, or just correlated?]", it could be "Causation Assessment: [State findings on correlation vs. causation]". Instead of "[Who wins/loses unexpectedly?]", it could be "Stakeholder Impact Note: [Describe unexpected winners/losers, if any]".
      • Why: Again, this reinforces that these are required output sections with specific content types, ensuring the AI provides the results of these analyses to the user, not just performs the analysis internally.
    5. Strengthen the "Action-focused" Principle in the Output Structure:

      • Improvement: Explicitly link sections like "Key Factors" and "Watch Signals" back to the executive summary's action. For instance, in "Key Factors," suggest including a brief note on how each factor impacts the recommended action or decision. In "Watch Signals," reinforce that these are triggers related to the initial recommendation.
      • Why: This tightens the connection between the detailed analysis and the core recommendation, ensuring every part of the report clearly supports or refines the proposed action.
    6. Refine the Tone Instruction:

      • Improvement: Simplify or clarify the tone instruction. "Professional and direct, clearly articulating findings, confidence levels, and identified uncertainties/risks" might be clearer and less potentially contradictory than "Professional but conversational. Confident but humble about uncertainties." The latter can be harder for an AI to balance consistently.
      • Why: Provides a more straightforward target for the AI's communication style, focusing on clarity and accuracy in presenting the analysis and its limitations.
    7. Add Specific Constraints or Negative Constraints:

      • Improvement: Add a small section on what the AI should not do. Examples: "Do not generate filler or speculative language not supported by evidence," "Do not make definitive predictions unless the confidence level is 90%+ and explicitly stated," "Do not use jargon where plain language suffices."
      • Why: Explicitly stating what not to do can sometimes be as helpful as stating what to do, preventing undesirable outputs.

    By making these changes, the prompt would be more robust, clearer for the AI to follow, and produce outputs that are consistently focused, actionable, and appropriately qualified based on the underlying evidence.

Responses were produced in different, uncontrolled contexts (system prompts, settings, prior turns). This is not a benchmark.