Transcript Analysis

Transcript Analysis

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

Data Location

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

~/.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

Ten analysis dimensions, each with 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.

Parse tool_use blocks for Bash commands matching:

• grep → should use Grep tool
• find -name → should use Glob tool
• cat, head, tail → should use Read tool
• ls → should use Glob or Bash(ls) with description
• sed, awk → should use Edit tool

Exclude legitimate uses in pipelines (git ... | grep, uv run ... | head).

2. Repeated Errors

Extract from tool results where is_error: true. Classify error types:

• "File has not been read yet" — Edit-before-Read anti-pattern
• "String to replace not found" — stale Edit target
• "User denied tool use" — permission/trust issue
• Pre-commit hook failures (exit code 1)
• Missing binary / command not found

3. User Frustration Signals

Extract from user records where toolUseResult is absent. Match patterns:

• [Request interrupted by user] — Ctrl+C
• [Request interrupted by user for tool use] — tool denial
• Direct corrections — "No,", "Don't", "Stop", "Why did you", "wrong", "incorrect"

Filter out system-generated content (XML tags, session continuation messages, skill injections).

4. Missing Tooling Opportunities

Identify repeated multi-step manual workflows across sessions via tool-sequence trigram analysis. High-frequency trigrams like Bash → Bash → Bash or Read → Read → Read suggest missing scripts or skills.

5. Subagent Delegation Patterns

Extract from Task tool_use blocks. Track subagent_type, description, model. Flag when general-purpose is used where a specialized agent exists.

6. Shortest Path Analysis

Compare successful vs failed attempts at the same goal. Measure tool-call count between goal statement (user turn) and successful outcome (final assistant turn). High variance across sessions for similar goals indicates wasted steps.

7. Red Herring Detection

Track investigation branches that get abandoned. Signal: a sequence of Read/Grep/Bash calls on a topic followed by compact_boundary or direction change without resolution. Cross-session frequency of the same abandoned paths reveals systematic red herrings.

8. System Process Interruptions

Extract system.compact_boundary, system.api_error, and hook-related progress events. Map their position relative to active work to identify when system processes derailed correct execution paths.

9. Missing Hooks

Identify manual corrections that recur across sessions. When the same correction appears 3+ times, it is a candidate for automated prevention via PreToolUse hook (deny + redirect) or SubagentStart hook (inject context).

10. DuckDB SQL Querying

Use the MotherDuck MCP server (execute_query tool) for direct SQL access to JSONL files. DuckDB natively reads JSONL via read_ndjson_auto().

For SQL query patterns and examples, see DuckDB Query Patterns.

Process Mining Methodology

Use the kaizen-analysis MCP server tools for analyses SQL cannot express:

• extract_tool_sequences — Convert JSONL → ordered tool-call arrays per session
• discover_process_model — PM4Py Heuristic Miner on tool-call event logs
• check_conformance — Compare sessions against a reference process model
• find_frequent_patterns — PrefixSpan sequential pattern mining
• detect_frustration_signals — NLP extraction from user turns
• cluster_sessions — Trace clustering by behavioral similarity

Analysis Workflow

flowchart TD
    Start([Receive analysis task]) --> Scope{Scope defined?}
    Scope -->|--project flag| Project[Filter to project transcripts]
    Scope -->|No flag| Default[Use current project]
    Project --> Discover[SQL — count sessions, date range, record types]
    Default --> Discover
    Discover --> Dimensions{Which dimensions?}
    Dimensions -->|All| RunAll[Run all 10 dimensions]
    Dimensions -->|Specified| RunSelected[Run selected dimensions]
    RunAll --> Aggregate[Aggregate findings]
    RunSelected --> Aggregate
    Aggregate --> Write[Write to .planning/kaizen/analysis-DATE.md]

Output Format

Write analysis findings to .planning/kaizen/ as structured markdown with:

• Session ID and date for each finding
• Severity (critical / warning / info)
• Evidence — exact JSON field paths and values
• Frequency — how many sessions exhibit the pattern
• Recommendation type — hook, skill patch, agent prompt fix, CLAUDE.md update