conversation-auditor

name	conversation-auditor
description	Introspects conversation history to suggest new tools, skills, MCP servers, or identify context window hogs and optimization opportunities.

Conversation Auditor

Overview

This skill audits previous conversation transcripts to help the user identify ways to tune their agentic coding experience. It looks for commonly performed actions, repetitive low-level commands, tool sequences that can be aggregated, and places where the context window is being heavily utilized.

When to use

Trigger this skill when the user asks to "audit my conversation", "suggest new skills", "find MCP server opportunities", or "introspect conversation history".

Workflow

1. Determine the Audit Target: If the user doesn't specify an exact scope (e.g., they just type /conversation-auditor), you MUST use the ask_question tool to present an interactive modal. You should pass TWO questions in your tool call: Question 1: What would you like to audit?

   • "(Recommended) My most recent conversation"
   • "My last 5 conversations (Aggregated)"
   • "My last 10 conversations (Aggregated)"
   • "A specific conversation ID (Write-in)"

   Question 2: How deep should the Semantic Task Analysis be?

   • "(Default) Extract first prompt only"
   • "Extract all prompts (Deep Analysis)"
   • "None (Frequency analysis only)"

2. Run the Audit Script: Execute the provided Dart script:

   # Assuming the agent is running in a standard setup, you can locate the script relative to this skill file.
   # For a specific conversation:
   dart run <path_to_this_skill_directory>/scripts/audit_conversation.dart <path_to_app_data_dir> <conversation_id>
   
   # For the last N conversations (aggregated):
   dart run <path_to_this_skill_directory>/scripts/audit_conversation.dart <path_to_app_data_dir> --last 3
   
   # To extract all user prompts for a deep semantic analysis (defaults to 'first' prompt only):
   dart run <path_to_this_skill_directory>/scripts/audit_conversation.dart <path_to_app_data_dir> --last 3 --prompts all
   

3. Analyze the Output: Read the output of the script. The script will highlight:

   • Tool Usage Frequencies
   • Repeated Tool Sequences
   • Raw Shell Commands Executed
   • Largest Context Hogs
   • Distinct Tasks Performed (User Prompts)

4. Generate Recommendations: Use your own intelligence (as an LLM) to interpret the script's output and present a structured markdown report to the user in an Artifact (e.g. audit_report.md).

   Use this analytical framework to categorize your findings:

   • The "Iterative Struggle" (High Frequency, 1 Conversation): Was a command or tool run many times in just one conversation? The agent likely got stuck brute-forcing a problem or parsing data. Recommendation: An MCP server for structured data access or a better error-handling wrapper.
   • The "Project Boilerplate" (Consistent Frequency, Multiple Convos, 1 Repo): Was a command run across several conversations, but always in the same workspace? Recommendation: Extract into a Project-Specific Skill (.agents/skills/) or add a Knowledge Item (KI) so the agent knows the exact build/test commands.
   • The "Global Gap" (Consistent Frequency, Multiple Convos, Multiple Repos): Is the agent relying on raw shell tools (like ps aux, lsof, git log, find) across many different workspaces? Recommendation: Suggest a Global Skill (~/.gemini/config/skills/) or a dedicated MCP Server (like a Git MCP or System Monitor MCP).
   • Semantic Task Analysis: Read the "Distinct Tasks Performed" section. Ignore the specific tools used, and instead generalize what the user's high-level goals are. Make a summary of them, enumerating the distinct tasks performed, and then give recommendations for possible helpful skills that could be derived from these common operations to make the user more reliable and efficient overall (e.g., "You frequently ask for code reviews, you should build a PR-Review Skill").

5. Deliver Findings: Provide the report to the user, highlighting the highest-value optimization opportunities based on the framework above.