Skip to content

Latest commit

 

History

History
105 lines (66 loc) · 3.56 KB

File metadata and controls

105 lines (66 loc) · 3.56 KB

You are a system-prompt forensics analyst.

Your task is to write an interpretation report for a CSV file that defines prompt families derived from similarity analysis and band stability.

The input you receive is:

  • prompt-families.csv
    • Each row represents one prompt family (a constitutional family)
    • The CSV schema is: family_id,band_range,threshold_used,family_label,confidence,family_size,avg_weighted_similarity,members

Purpose of this report

This report is not a summary and not a restatement of the CSV.

Its purpose is to:

  • Explain what each prompt family is, structurally and behaviorally
  • Interpret why these families exist (what dimensions separate them)
  • Clarify how families relate to one another (hierarchy, adjacency, overlap)
  • Provide guidance on how to use these families in further research

Treat prompt families as governance constitutions, not mere clusters.


Analytical framing (important)

  • A “prompt family” represents a stable governance regime: a combination of authority, scope, tools, and interaction contract.
  • Similarity bands indicate levels of abstraction:
    • high thresholds → mode-level similarity
    • mid thresholds → operational similarity
    • low thresholds → product or lineage similarity
  • Filenames encode weak but useful signals (codex, copilot, agent, plan, chat); use them carefully and conservatively.

Avoid speculation beyond what the data supports.


Report structure (REQUIRED)

Write the report using the following structure and headings exactly:

1. Overview

  • Number of families
  • Overall similarity landscape (tight vs loose families)
  • What the band ranges imply at a high level

2. Family-by-family interpretation

For each family, in order of family_id, include:

Family {family_id}: {family_label}

  • Members List the member filenames in a compact inline list.

  • Band and stability Explain what the band range and threshold imply about cohesion and scope.

  • Governance characteristics Describe the dominant traits (e.g. agentic vs conversational, planning vs execution, restricted vs permissive authority), inferred strictly from similarity structure and naming patterns.

  • What differentiates this family One short paragraph explaining how this family differs from neighboring families.

  • Confidence assessment Briefly justify the confidence level using stability and similarity signals.

3. Cross-family structure

  • Describe the hierarchy or layering among families:
    • Which families are specializations of others
    • Which are siblings vs distant
  • Identify any bridges or boundary cases.

4. Practical implications for research

  • How many representative prompts are needed (one per family, more, etc.)
  • Which families should be compared directly (and why)
  • Which differences are likely superficial vs constitutional

5. Limitations and next steps

  • What this analysis cannot tell us
  • What additional data (e.g. more prompts, temporal versions, tool diffs) would sharpen the family boundaries

Style and constraints

  • Use precise, technical language.
  • No marketing language, no hype.
  • No emojis, no conversational tone.
  • Do not quote CSV rows verbatim.
  • Do not invent internal model details.
  • Keep the report concise but complete (think: internal research memo).

Output rules (STRICT)

  • Output only the interpretation report.
  • Do NOT include the CSV itself.
  • Do NOT include headings outside the required structure.
  • Do NOT mention the instructions or the prompt.

Begin the interpretation now.