Hatena
A Full-Picture Guide to the Recommended Claude Code Setup: How Harness, Skills, and MCP Work, with Examples
AIFirst seen 393 points
A Japanese writer explains the AI setup on his computer, built around Claude Code. Every request goes through a "harness" of checks and permission rules before it is sent to other AIs, skills, tools, or outside services.
One entry point, one control center
The writer uses Claude Code as the only place he gives orders. Before any work starts, the request passes through a harness. The harness checks what the request is, whether the action is allowed, and whether it is safe. Then the request is sent to the right helper, such as another AI, a skill, a tool connector (MCP), or an outside service. Small helper AIs called subagents do the work and check it. The final result comes out of the last layer.
Key numbers from the setup
- Setup as of November 2026: Claude Opus 5.5 model, 147 skills, 150 subagents, and 17 plugins.
- There are 4 ways to reach Claude Code: a phone app, a terminal, Discord, and scheduled tasks.
- The harness has 5 parts: a judge AI called Jev, 11 hooks, permission rules in settings.json, API keys in one .env file, and a CLAUDE.md file.
- Jev answers in about 0.2 to 0.35 seconds.
- The .env file holds 27 API keys as of November 2026.
What the harness is
The harness is the set of checks that every request must pass. Judge AIs read text or images and give a score or a yes-or-no answer. Hooks start set tasks automatically when something happens. Permission rules say what Claude Code may do, what needs a confirmation, and what is forbidden. API keys are passwords for outside tools, and the writer keeps them in one file. The writer stores the keys safely with 1Password, a password manager.
Why this matters
Many people think bad results come from poor instructions. This article argues that the setup around the AI matters just as much. A harness that checks every step can make results more steady in quality. It can also block risky actions before they cause harm. For a student, the key idea is simple: a good AI tool works best when it has clear rules, checks, and a clear path from request to result.
What to watch
- The Discord bot called Kuro was stopped on November 5, so its future use is unclear.
- The Grok Bot is limited by its usage cap, so the writer moved most work to Claude Code on a cloud PC.
- The writer plans to keep building the knowledge layer, since the article says better knowledge improves every layer above it.
Glossary
- Claude Code
- An AI coding and work assistant from Anthropic that runs commands and edits files on a computer.
- Jev
- A judge AI from TypeSafe that reads text and returns scores or decisions, but does not write text.
- MCP
- A standard way for AI tools to connect to outside apps and data sources.
- settings.json
- A settings file that lists what Claude Code may do, what needs confirmation, and what is forbidden.
- 1Password
- A password manager that stores secrets such as API keys in a safe vault.