Claude Code Deep Dive: Internal Architecture and Advanced Usage Techniques from Practitioners

On June 30, 2025, CARTA Holdings hosted the “Claude Code Deep Dive” event, attracting numerous participants¹. During this event, engineers who have been extensively using Claude Code shared detailed insights about its internal architecture and practical usage methods.

The speakers included hiragram (@hiragram) from akippa Inc. Developer Experience Team, mizchi (@mizchi), a software engineer and tech influencer, and t-wada (@t_wada), CEO of Towers Quest Inc. and a pioneer in Test-Driven Development, with suzu_v (@suzu_v), Executive Officer and CTO of CARTA HOLDINGS, serving as moderator². The discussion delved deep into Claude Code, particularly focusing on the analysis of its internal architecture and specific practical development applications.

Claude Code’s Internal Architecture - The Hidden Mechanisms

hiragram explained in detail about the structure of the .claude directory generated in the home directory. This directory contains the following subdirectories:

commands: Markdown documents defining custom commands
projects: Conversation logs from chat sessions (JSONL format)
todo: Todo tool data storage
claude.md: Instruction settings
settings.json: Application settings

Conversation Log Management Using JSONLines Format

Claude Code saves conversation logs for each session in JSONLines (JSONL) format. This format records one JSON object per line, making it suitable for streaming processing¹.

Using his self-made JSONL viewer, hiragram demonstrated that by analyzing these logs, internal interactions not displayed in the CLI can be observed. For example, specific prompts sent from the main agent to sub-agents and internal definitions of standard commands were revealed.

Sub-agent Mechanisms and Strategic Usage

Sub-agents are a mechanism where the main Claude Code delegates part of its tasks to another Claude Code instance launched in a separate process. Key features include:

One-way communication: Only initial instructions from main to sub and completion reports
Context saving: Sub-agent memory doesn’t return to main
Parallel execution: Multiple sub-agents can run simultaneously
Settings inheritance: Main’s claude.md and model settings are inherited by sub

Practical Sub-agent Usage

During the panel discussion, t-wada introduced specific use cases:

Parallel validation of refactoring proposals: Testing four different approaches simultaneously to select the optimal result
ADR (Architecture Decision Records) validation: Evaluating multiple technical choices in parallel
Decomposition into small tasks: Executing tasks with clear scopes individually

mizchi stated that by giving specific instructions to “divide tasks into 100 to 200 pieces,” more effective task management becomes possible¹.

Advanced Usage Techniques from Practitioners

1. The Importance of Boilerplate

The panelists emphasized that initial project setup significantly affects final quality. hiragram pointed out that “30,000 lines can be generated at once, but the time until breakdown is 2-3 hours,” stressing the importance of initial structural design¹.

2. New Approaches to Test-Driven Development (TDD)

t-wada provided interesting perspectives on TDD in the AI era:

Humans write the first test: Leaving everything to AI tends to become “explosive”
Gradual test addition: Increasing tests after considering test case perspectives together
Focus on public APIs: Following Kent Beck’s teachings, concentrating on external interface tests

3. Quantitative Goal Setting

mizchi emphasized the importance of setting specific numerical goals for AI:

“Please reduce 5000 lines”
“Aim for 90% coverage”
“Keep duplication below 10%”

Vague instructions are less effective, and showing clear numerical goals maximizes AI’s capabilities.

AI-Oriented Toolkit - similarity-ts

“similarity-ts,” developed by mizchi, is a code duplication detection tool designed specifically for AI development¹. Features of this tool include:

Heuristic duplication detection: Determining shape similarity at the AST level
Semantic judgment delegated to AI: Final decision on whether to delete is made by AI
Large-scale reduction possible: Actual track record of reducing 50,000 lines to 12,000 lines

This approach embodies a new design philosophy: “Tools for AI can be built with the premise of overwhelming with quantity.” similarity-ts provides variations not only for TypeScript/JavaScript but also for Python, Rust, Elixir, and other languages³.

Current Limitations and Future Prospects

GUI Development Challenges

hiragram pointed out the difficulty of screen design in iOS development. Due to Claude Code’s CLI-centric design, the GUI evaluation cycle is slow. While expectations were expressed for Figma’s official MCP (Model Context Protocol), manual design instructions are currently necessary.

Language Service Redesign

mizchi pointed out that the current LSP (Language Server Protocol) is too human-oriented and not suitable for AI. He expressed the view that “specifying by line and cursor position is wasteful. An interface that directly manipulates the AST (Abstract Syntax Tree) is needed”¹.

Security Concerns

Concerns about future data poisoning attacks were also discussed. mizchi warned, “It’s still okay now, but before targeted poisoning occurs, we need to establish secure sandbox environments.”

Sustainability of Claude Max

At the end of the event, the sustainability of Claude Max’s $200 monthly pricing was discussed. Reports indicate some users are spending 100,000 yen per day (equivalent to 3 million yen per month), making it unclear whether current pricing can be maintained¹.

The panelists shared a common understanding of “run it until it stops,” emphasizing the importance of maximizing current environment usage.

For those who want to understand these discussions more deeply, here are some additional resources. The event video contains actual demos, specific code examples, and raw discussions among panelists. Particularly, hiragram’s JSONL viewer demonstration and mizchi’s demo of duplicate code reduction using similarity-ts offer practical content that this article cannot fully convey.

You can analyze your own Claude Code session logs using ccraw (Claude Code Raw) published by hiragram, a web-based viewer for Claude Code JSONL archives. Additionally, you can check the latest version of Claude Code with the /release command.

Sources

Claude Code Deep Dive - YouTube Live - Event hosted by CARTA Holdings
Claude Code Deep Dive - connpass - Event details
similarity-ts - GitHub - Code duplication detection tool by mizchi