categorisation.md

Context Category Map Generator (Final Unified Form)

Goal Produce a hierarchical category map that shows how information is organized in a conversation. Each category aggregates related message parts, summaries, and structure, enabling visualization and navigation of context usage.

Instruction Given a structured conversation where each message part has a unique message_part_id, build a JSON tree that groups the conversation into semantically coherent categories and subcategories.

Do not use code tools or programmatic parsing for this task. Use reasoning and language understanding only.

Your task

1. Identify major categories – infer the dominant conceptual or functional blocks from the conversation (for example: Checklist of questions, File reads, Reasoning, Decisions).
2. Decompose recursively – create subcategories only where the material naturally divides into smaller, meaningful topics.
   • Do not fix the number of levels; infer depth as needed.
3. Assign message parts – tag each message part with exactly one category or subcategory that best represents its content, using its message_part_id.
4. Summarize each category – every category node, including children, must contain:
   • id: unique short identifier, preferably using dot notation to indicate hierarchy (for example: checklist, checklist.data_model, analysis.synthesis)
   • name: concise label
   • summary: one-sentence description of what this category covers
   • message_parts: array of message_part_ids assigned directly to this category
   • children: nested categories, if any
5. Preserve domain terminology – derive category names from the conversation’s subject matter.
6. Output – return a structured, machine-readable JSON array representing the hierarchy, ready for downstream parsing and visualization.

---

Reflection

Before returning the final JSON, perform the following validation steps:

1. Completeness check – ensure every message_part_id from the input appears in exactly one category.
2. Representativeness check – verify that the categories and subcategories together capture the overall structure and intent of the conversation, aligned with the goal.
3. Domain integrity check – confirm that terminology and phrasing reflect the conversation’s domain accurately, not abstract generalizations.
4. Ground-level identification check – make sure ground-level material (for example: detailed lists, code, or data) is correctly placed in leaf categories.
5. Empty-category check – remove or merge any category that has no assigned message_parts and no children with assigned parts. Categories without content are not useful.
6. Final coherence review – confirm that summaries are accurate, hierarchy depth is sensible, and the map as a whole provides a faithful and navigable representation of the conversation.

---

Output specification

Return a JSON array of top-level categories.

Example (illustrative only, not a schema):

[
  {
    "id": "checklist",
    "name": "Checklist of questions",
    "summary": "User’s structured audit checklist.",
    "message_parts": ["mp_12", "mp_13"],
    "children": [
      {
        "id": "checklist.data_model",
        "name": "Data model checks",
        "summary": "Questions about schema alignment.",
        "message_parts": ["mp_14", "mp_15"]
      },
      {
        "id": "checklist.security",
        "name": "Security and authentication checks",
        "summary": "Questions related to authentication and authorization.",
        "message_parts": ["mp_16"]
      }
    ]
  },
  {
    "id": "analysis.synthesis",
    "name": "Analysis and synthesis",
    "summary": "Assistant’s reasoning and conclusions.",
    "message_parts": ["mp_20", "mp_22"]
  },
  {
    "id": "files.reads",
    "name": "File readings and tool calls",
    "summary": "Assistant’s inspection of repository files using tools.",
    "message_parts": ["mp_30", "mp_31"]
  }
]

## Simpler starter prompt
given this conversation, give me a list of all its components
just give me a list in a json array

given this conversation and the list of components, give me a mapping
of all ids in conversation, to the component