Claude Code advanced skills: 393 points HN 7 core practices of explosive articles

Core conclusion

On May 26, 2026, developer Arpan Patel published an in-depth guide titled "Beyond the Prompt: Claude Code". The system dissected all the configuration and workflow techniques required for Claude Code from casual user to daily driver. The article received 393 points and 237 comments on Hacker News, making it one of the most talked about AI coding tools topics this week.

Key Points

• Time of incident: 2026-05-26 -Affected objects: Claude Code users, AI coding Agent users, automation practitioners
• Core changes: A complete practical guide reveals the huge gap between casual users and power users, and exposes practical problems such as the fragmentation of the skill system in the current AI coding Agent ecosystem and the questionable cost-effectiveness of LSP plug-ins.

Background: from prompt to programmable agent

The author of the article, Arpan Patel, made a sharp observation at the beginning: casual users regard Claude Code as "fancier autocomplete", while daily driver regards it as "a programmable agent with memory, custom commands, parallel sessions, and a project setup that compounds over time". The efficiency gap between these two usage methods does not grow linearly, but exponentially.

At the heart of this difference is a cognitive shift: once you think of Claude Code as an autonomous agent that requires guardrails, rather than a prompt-and-wait chatbot, the entire workflow changes qualitatively.

Detailed explanation of 7 core skills

1. Plan Mode as a design document

The core principle comes from Boris Cherny of the Anthropic team: Give Claude a way to validate his work. Without a verification loop, you are the only feedback loop; with it, Claude can iterate autonomously until the code actually works. Boris claims this approach results in a 2-3x improvement in quality.

Specific operations: Press Shift+Tab twice to enter Plan Mode, and Claude enters read-only exploration mode. Read the file, trace the data flow, understand the data model, and then output the plan. Finally review the plan with a separate Claude session (as Staff Engineer) - it's easier to spot blind spots because there's no contextual bias.

2. Delegation pattern instead of pair programming

Tip from Claude Code team member Cat Wu: Think of the model as an engineer to whom you delegate tasks, rather than a pair-programming partner to whom you guide line-by-line. Write a clear brief and then let the model execute independently.

Two practical shortcut keys: Ctrl+G to open Claude's plan and modify it directly in the editor; when Claude makes a mistake, end with "Update CLAUDE.md so you do not repeat this" - Boris calls it the most compounding habit.

3. Hierarchical configuration system of .claude directory

Most users open the .claude/ directory, see CLAUDE.md and don't go any further. But the actual configuration system is much more than that:

Core mental model: Project files describe the project, global files describe yourself.

4. Skills, Subagents, Plugins system

Claude Code provides three mechanisms to implement reusable workflows: Skills are reusable prompts in Markdown format (called via /name), Subagents are independent Agents that can run in parallel in the background, and Plugins are a collection of pre-installed skills and subagents.

However, a user in the HN comment area pointed out that the current system has a serious fragmentation problem: if you want to do a code review, there are five different ways to complete it, but there is no clear guide to tell users which method to choose in which scenario.

5. MCP tool integration

.mcp.json allows teams to share MCP server configuration. This is the standard way to expose external tools (database, API, file system) directly to Claude Code. The MCP protocol allows AI Agents to call any external tools, thereby pushing the boundaries of coding itself.

6. Self-verification closed loop is the key

Giving Claude a way to validate his work is the most important tip of all. Type checking, test running, linting commands - pass the verification steps as part of the prompt to Claude and it will automatically verify and fix it after every change. This closed-loop mechanism is the source of 2-3x quality improvements.

7. Recursive CLAUDE.md evolution

The most compounding habit: every time Claude makes a mistake, ask it to update CLAUDE.md. Claude is surprisingly good at writing rules for his own mistakes. Over the course of a few months, your CLAUDE.md will accumulate a set of highly customized project specifications that are automatically loaded every session.

A critical addition to the HN community

Of the 237 comments, the most notable discussion focused on the LSP plugin issue. One user shared his experiment: installed the LSP plugin for Rust, Python and Dart, and checked the session log two months later - the agent only called the LSP tool once. Claude Code has done the same thing himself using ripgrep, cargo clippy, dart analyze, etc. LSP's RAM consumption, especially when accumulated across multiple language projects, creates a performance burden.

Another user pointed out that although there are many ways to expand the current Claude Code ecosystem, there is a lack of clear best practice guidelines to guide users in making choices.

Implications for AI automation practitioners

The value of this guide goes beyond Claude Code itself. It reveals a broader trend: AI-coded agents are evolving from question-and-answer toys into carefully engineered productivity systems.

For WayToClawEarn readers, the following three practical points are worth implementing immediately:

• Create a CLAUDE.md for your project, including project constraints, coding standards and verification steps
• Use Plan Mode as the design document, and the second Claude does the review to form a closed loop
• After each bug fix, ask Claude to update CLAUDE.md - this compound interest effect will be visible after a month

Tool entry (trigger tool floating card)

AI tools that appear naturally in the text: Claude Code, OpenAI, ChatGPT, Anthropic, n8n, Hermes Agent, DeepSeek

Internal link guidance

• Want a cheaper solution to run Claude Code? Watch: How to use DeepSeek V4 instead of Anthropic to run Claude Code (https://waytoclawearn.com/tutorials/guide-claude-code-deepseek-v4-cheap-alternative)
• Don’t know where to start when choosing an AI coding agent? Watch: AI Programming Agent Technology Selection Three-Dimensional Decision-making Framework (https://waytoclawearn.com/tutorials/guide-ai-coding-agent-tech-stack-selection-strategy)
• Someone used Claude Code to start a business in 48 hours: a real case of achieving a monthly income of $9,000 in 3 months (https://waytoclawearn.com/cases/claude-code-9000-mrr-thinkidiot-48hrs)
• Claude Code + AWS build AI SaaS: practical review of earning $12,000 in 3 months (https://waytoclawearn.com/cases/claude-code-platform-aws-saas-12000-month)