conversation-summarizer

Summarize AI text conversations into structured, multi-level overviews.

Workflow

1. Detect the conversation format
2. Parse the conversation into a normalized turn sequence
3. Generate the requested summary level(s)
4. Present the summary

Step 1: Detect and Parse the Conversation Format

Identify the format by examining the input structure. See references/format-detection.md for format signatures and parsing guidance.

Normalized turn structure (internal representation after parsing):

Each turn should resolve to:

• role: "user", "assistant", or "system"
• content: The substantive text of the turn
• turn_number: Sequential position in conversation
• metadata: Any additional fields (timestamps, actions, tool calls)

If the format is unrecognized, inform the user and ask for clarification about the structure. Do not guess.

Step 2: Generate Summary

Determine which summary levels to produce based on the user's request. If the user does not specify, default to producing all applicable levels.

Level 1: High-Level Overview

Produce a single paragraph (3-5 sentences) capturing:

• What the conversation was about (topic/goal)
• The main outcome or conclusion
• The overall arc (e.g., "exploration -> decision -> implementation")

Level 2: Key Decisions and Outcomes

Extract a bulleted list of decisions made, conclusions reached, or significant turning points. For each decision:

• State the decision clearly
• Note which turn(s) it emerged from
• Note any alternatives that were considered and rejected

Level 3: Action Items and Follow-ups

Extract concrete next steps, tasks, or commitments. For each:

• State the action item
• Note who is responsible (user or assistant) if discernible
• Note any deadlines or conditions mentioned

Level 4: Per-Turn Summary

Produce a sequential summary with one entry per meaningful turn:

• Skip turns that are purely procedural (e.g., "OK", "thanks")
• Consolidate adjacent turns from the same role when they form a single logical unit
• Keep each entry to 1-2 sentences

Handling Edge Cases

• Very long conversations (50+ turns): Group turns into phases or topics before summarizing. Identify natural breakpoints where the conversation shifted focus.
• Multi-topic conversations: Organize summaries by topic rather than strictly chronologically. Note where topics interleave.
• Conversations with code/technical content: Summarize the intent and outcome of code blocks rather than reproducing code. Note file paths, function names, and key technical decisions.
• Conversations with tool use/actions: Summarize tool invocations by their purpose and result, not by the raw tool call syntax.
• Incomplete or truncated conversations: Note that the conversation appears truncated and summarize only what is present.

Output Format

Present summaries in clean markdown. Use this structure, adapting sections based on what the user requested:

# Conversation Summary

## Overview
[Level 1 content]

## Key Decisions
[Level 2 content]

## Action Items
[Level 3 content]

## Turn-by-Turn Summary
[Level 4 content]

If the user requests only specific levels, produce only those sections. If the conversation metadata includes a date or project name, include it in the header.