Claude Code Daily Driver: Skills, Subagents, Plugins & MCPs

Arpan Patel on Advanced Claude Code Usage

Arpan Patel published a detailed guide on Hacker News in late 2025 covering Claude Code configuration and extension. The post draws from Boris Cherny and the Anthropic team to explain how to move beyond basic prompting. It focuses on project-level files, custom behaviors, and parallel sessions that compound across work. Patel emphasizes verification loops and delegation patterns that reduce manual oversight.

CLAUDE.md Setup and Skills

The .claude directory stores persistent instructions and extensions for a project. CLAUDE.md sits at the root and defines conventions, testing commands, and review criteria that Claude loads on every session. Patel includes examples from the Claude Code team that list exact file paths, lint rules, and commit message formats rather than vague goals.

CLAUDE.local.md overrides the shared file for machine-specific settings such as local database URLs or personal shortcuts. Skills extend this further by packaging reusable instructions into named commands. A skill for Go API conventions might contain rules for handler structure, error wrapping, and test file placement. Once installed, typing the skill name injects the full set of constraints into the current context without repeated prompting.

Subagents, Plugins, and Commands

Subagents run as separate Claude instances with their own system prompts. A /pr-review subagent can be configured to check diff size, test coverage, and security patterns before any merge. Patel lists several community subagents that handle tasks such as dependency updates and migration scaffolding.

Plugins add marketplace extensions for common workflows. Underused built-in commands include /goal, which forces Claude to state success criteria before implementation begins. Plan mode, activated with Shift+Tab twice, restricts Claude to read-only exploration so it can trace data flows without editing files. Patel recommends running one session in plan mode to produce a design document, then opening a second clean session to critique that document for gaps.

MCPs and Daily Workflow

MCPs connect Claude Code to external tools such as Obsidian vaults or local databases. A typical Obsidian workflow lets Claude read and write notes while maintaining links and tags according to a defined schema. Patel notes that these connections work best when paired with explicit verification steps inside the MCP configuration itself.

The core pattern across these features is delegation over line-by-line guidance. Reference files with @ paths or pipe error output directly rather than describing symptoms. When implementation stalls, return to plan mode to insert new verification checkpoints. These steps add setup time upfront but reduce context switching and rework in longer sessions.

---

📖 Related articles

• Agentic Coding: Una Trappola per lo Sviluppo Software?
• Lean-ctx: Ottimizzatore Ibrido Riduce Consumo Token LLM del 89-99%
• Rust rivoluziona Claude Code: Avvio 2.5x più rapido e volume ridotto del 97%