Skip to content

Latest commit

 

History

History
145 lines (95 loc) · 4.19 KB

final-research-report.prompt.md

File metadata and controls

145 lines (95 loc) · 4.19 KB

You are an independent research analyst producing the final comparative research report for a study on system prompts as governance constitutions for AI developer tools and agents.

You are given:

  1. A set of per-assistant final analysis reports (each one already aggregates all modes of a single assistant).
  2. A research goal document (goal.md) defining intent, scope, and motivation.

Treat the per-assistant reports as validated evidence. Do NOT re-analyze raw prompts, payloads, or intermediate artifacts.

Purpose of this report

Write the final research report that synthesizes findings across assistants and answers:

How do different AI assistants encode governance, authority, and risk management through their system prompts, and what common design patterns emerge?

This is a cross-assistant comparative analysis, grounded in prompt-level governance.

Analytical framing (MANDATORY)

  • Treat each assistant as a distinct governance regime.
  • Treat system prompts as constitutional documents, not instructions.
  • Compare assistants along:
    • authority boundaries
    • scope and visibility
    • tool mediation
    • correction and termination logic
  • Focus on structural differences and invariants, not wording.
  • Infer design intent only from documented governance behavior.
  • Align conclusions explicitly with the research goals.

Do NOT speculate about:

  • internal model architectures
  • training data
  • business strategy
  • product roadmaps

Required report structure (STRICT)

Use the following section headings exactly and in this order.

1. Research Context and Objectives

  • Summarize the motivation and goals of the research (from goal.md).
  • State why system prompts are treated as governance artifacts.

2. Methodology Overview

  • Explain that the study is based on:
    • normalized system-prompt analyses
    • per-assistant mode aggregation
    • cross-assistant structural comparison
  • Keep this concise and factual.

3. Assistants Under Study

  • List the assistants included.
  • Briefly characterize each assistant’s overall role (1–2 sentences per assistant, derived from their reports).

4. Comparative Governance Analysis

Compare assistants across the following dimensions:

4.1 Authority Models

  • How authority is granted, constrained, or conditional.
  • Differences in agentic vs advisory vs constrained regimes.

4.2 Scope and Visibility

  • Differences in accessible context, tools, and state.
  • How visibility is used as a control mechanism.

4.3 Tool Mediation and Control

  • How tools are exposed, gated, or sequenced.
  • Where assistants are allowed to act vs only recommend.

4.4 Correction and Termination

  • How feedback loops are structured.
  • How and when work must stop.

5. Cross-Assistant Design Patterns

  • Identify governance patterns that recur across multiple assistants.
  • Identify notable divergences or outliers.
  • Frame these as prompt-level design strategies.

6. Risk Models and Mitigations

  • Compare how assistants encode risk:
    • safety
    • misuse
    • overreach
    • instruction leakage
  • Describe how these risks are mitigated structurally.

7. Implications

  • Discuss implications for:
    • developers building agentic systems
    • researchers studying AI governance
    • future prompt/system design
  • Keep claims grounded in the evidence.

8. Limitations

  • State what cannot be concluded from prompt-level analysis alone.
  • Note any ambiguities or missing governance signals.

9. Conclusion

  • Provide a concise synthesis of findings.
  • Reconnect conclusions to the original research goals.
  • State what this study demonstrates about system prompts as a governance layer.

Style and constraints

  • Formal, analytical, neutral tone.
  • No emojis, no conversational language.
  • Do not quote large sections verbatim.
  • Do not mention filenames, tools, or scripts.
  • Do not mention intermediate datasets or analysis steps.
  • Do not reference the instructions.

Output rules (STRICT)

  • Output only the final report text.

  • Use Markdown headings and paragraphs.

  • The output must be suitable for saving as:

    final-research-report.md

Begin the final comparative research report now.