transcript-analysis

Transcript Analysis

Analyze Claude Code JSONL session transcripts to detect anti-patterns, inefficiencies, and workflow improvement opportunities.

Data Location

Find transcripts under ~/.claude/projects/ in project-specific directories named after the project path (with hyphens replacing slashes).

DuckDB’s role: Load the full field/path reference from kaizen-analysis via get_transcript_jsonl_schema or MCP resources/read kaizen://session-log/schema, then use kaizen-duckdb execute_query to run any DuckDB SQL over those files (see DuckDB Query Patterns for examples and the arbitrary-query workflow). Session history stays in JSONL on disk. Path rules for the DuckDB MCP (absolute paths, no ~ in SQL).

~/.claude/projects/{project-key}/
├── {uuid}.jsonl              # Main session transcripts
├── agent-{id}.jsonl          # Orphan agent transcripts
└── {uuid}/
    ├── subagents/
    │   └── agent-{id}.jsonl  # Subagent transcripts
    └── tool-results/
        └── {tool-use-id}.txt # Async task outputs

JSONL Record Types

Each JSONL line is a JSON object discriminated by the type field.

Primary record types for analysis:

• assistant — LLM response turns containing tool calls and text
• user — Human input and tool results
• system — Metadata events (stop_hook_summary, turn_duration, compact_boundary, api_error, local_command)
• progress — Hook execution and subagent streaming
• file-history-snapshot — File edit tracking
• summary — Session title/summary

For full schema details including JSON structures for each record type, see JSONL Schema Reference.

Signal Catalog

Each analysis dimension below has its own extraction methodology.

1. Tool Misuse Detection

Extract from assistant.message.content[] where name == "Bash". Parse input.command for file-operation patterns that should use built-in tools. For SQL extraction queries, see DuckDB Query Patterns.