The Creator of Claude Code Shares 10 High-Efficiency AI Coding Tips

A Viral 28-Minute Talk

Boris, the creator of Claude Code, recently delivered a 28-minute technical talk that racked up 3.5 million views on X (formerly Twitter), quickly going viral. The core of the talk was his 10 practical tips that reportedly can double the effectiveness of using Claude.

Boris Cherny is the technical lead for Anthropic's Claude Code product, having spearheaded the entire journey from prototype to official release. Unlike traditional IDE plugins, Claude Code runs directly in the terminal environment, capable of reading and writing files, executing commands, and interacting with version control systems. Its core design philosophy is "agentic coding" — AI isn't just a code completion assistant, but a programming agent that can autonomously complete multi-step tasks.

Interestingly, the Claude team had previously published a batch of usage tips, but these are 10 entirely new ones. More importantly, these recommendations aren't limited to Claude itself — even if you're using other AI coding tools, you can benefit from them just the same. What they all have in common is extreme practicality — nearly all of them can be applied immediately.

This article breaks down the thinking behind these 10 tips to help you rethink how to collaborate effectively with AI coding tools.

Tip 1: Set Up Before You Start

Many people grab Claude Code and rush straight into coding, but Boris recommends spending a few minutes on configuration first. He states plainly: installing it without these settings is like not installing it at all.

Core settings include:

• Allowed Tools command: Explicitly tell Claude which tools it can invoke, preventing unnecessary operations. This is essentially a permission control mechanism — AI agents may attempt to run various system commands during task execution, and whitelist restrictions can prevent misoperations, which is especially critical in production environments;
• Terminal Setup: Enable Shift+Enter for line breaks, which greatly improves the experience when writing long prompts. Since Claude Code runs in a terminal environment, the default Enter key submits input directly, and multi-line editing requires additional configuration;
• Theme switching: Use the /theme command to adjust to light or dark mode. If you prefer the dark aesthetic, just use /theme dark — easy on the eyes and efficient;
• Install GitHub App: Once installed, you can trigger AI tasks directly in Issues and PRs by @Claude, dramatically improving team collaboration efficiency. This means team members don't need to switch to the terminal — they can involve AI directly within the code review workflow;
• Enable notifications: Get timely alerts when time-consuming tasks complete, so you never miss critical milestones. Since AI agents may need several minutes to complete complex tasks, the notification mechanism lets developers switch to other work during wait times;
• Mac dictation feature: Go to system settings to enable dictation, then double-tap the dictation key to describe requirements by voice, as naturally as talking to a person.

These settings may seem trivial, but they determine how smooth your subsequent experience will be. The more complete your setup, the clearer the AI tool's boundaries, and the lower the probability of errors.

Tip 2: Start by Asking AI Questions

Boris particularly emphasized a shift in mindset. While more and more developers are embracing web coding (programming in browser or AI environments), many people's thinking is still stuck in the traditional mode — encountering problems and trying to solve everything on their own.

He suggests that companies, when onboarding new hires, shouldn't pile on various tools right away or rush into coding. Instead, start with asking questions. Let new team members understand Claude's capability boundaries through questions: what it can do, and what it can't.

The deeper value of this advice is: in the AI coding era, knowing how to ask questions is more important than knowing how to write code. Only by understanding the tool's capability boundaries can you reasonably decompose complex tasks and efficiently hand them off to AI. This is analogous to the "requirements analysis" phase in software engineering — clearly describing the problem is often more critical than solving it. In the context of AI agents, a precise question can save the time cost of multiple interaction rounds, while vague instructions may lead AI down the wrong path, requiring a complete restart.

Tips 3 & 4: Plan First and Guide Execution

Before actually coding, proper planning is crucial. If you're not good at planning, you can directly ask AI to draft a plan for you, and after the first draft is complete, have AI self-review and suggest modifications. In other words, you can absolutely use AI as a product manager or project manager.

This "let AI self-review" approach is technically called "Reflection" or "Self-Critique" mode, and is one of the effective strategies for improving LLM output quality. By requiring the model to examine its first-round output in a second round of conversation, it often discovers logical gaps or missed edge cases.

Following this is guiding Claude to operate your way. A practical tip: first have AI propose several solutions to a problem, then you select one for it to execute. This effectively prevents AI from "going rogue" and deviating from expected directions.

The core of these two tips is: humans are responsible for decisions, AI is responsible for execution — clear division of labor maximizes efficiency. This is also the mainstream consensus in the current AI coding field — the "human-in-the-loop" working model, which leverages AI's speed and breadth advantages while preserving human judgment in architectural design and key decisions.

Tips 5-7: Connect Tools, Sync Workflows, Feed Context

Connect Team Tools and MCP

Connecting your code hosting tools and MCP (Model Context Protocol) tools to Claude makes subsequent operations smoother. MCP is a communication standard open-sourced by Anthropic in late 2024, designed to solve the connection problem between AI models and external tools/data sources. Before MCP, every AI tool needed custom integration code for different external services, leading to severe ecosystem fragmentation. MCP defines a unified client-server communication protocol that lets AI models call external capabilities like database queries, API requests, and file operations through standardized interfaces — similar to what the USB protocol does for hardware devices, providing a universal "port" that lets any protocol-compliant tool plug and play.

Sync Workflows

Boris outlined three common workflows. His most frequently used is the first — making a plan first when facing complex tasks; if encountering error-prone or ambiguous parts, he uses the second workflow: writing test cases first.

The "write test cases first" approach originates from TDD (Test-Driven Development) in software engineering. In AI coding scenarios, test cases not only verify the correctness of final code, but more importantly provide AI with clear "success criteria" — AI can repeatedly run tests for self-verification without needing manual line-by-line inspection of output results.

More Context Means Better AI Performance

You can preset core rules through CLAUDE.md files to constrain AI behavior. CLAUDE.md is a project configuration file specific to Claude Code, placed in the project root directory. AI automatically reads its contents as context at every startup. Developers can define code style specifications, project architecture descriptions, common commands, prohibited operations, and more within it. Claude Code supports multi-level configuration files: the root directory's CLAUDE.md serves as global rules, while CLAUDE.md files in subdirectories can override or supplement rules for specific modules, similar to the layered design philosophy of .gitignore or .editorconfig.

Boris recommends spending extra time organizing context information, as this directly improves output quality. Other information can be brought in using / commands or directly with @filename.

Context information essentially tells AI "what your project is and what you expect." The more complete the information, the more accurate AI's understanding, and the higher the output quality. This is directly related to how large language models work — models reason based on information within the context window, and the richer and more structured the context, the higher the reasoning quality.

Tip 8: Team-Shared Permissions, Memory & MCP

This step focuses on team collaboration scenarios. Boris shared his team's practices:

• Build a Prompt library: Good prompts are rare finds — save them and share across the entire team;
• Share CLAUDE.md and Memory.md documents: Achieve information alignment and synchronized execution standards;
• Share internally-developed MCP services: For example, the team developed an MCP service for fetching trending topics that the whole team can call. This point is considered "extremely critical."

The essence of sharing mechanisms is transforming individual experience into team assets, leveling up AI tool usage across the team and avoiding repeated mistakes. This aligns with the DevOps concept of "Infrastructure as Code" — making implicit operational experience explicit and codified so it can be version-controlled, reused, and continuously optimized. In the AI coding context, prompt libraries and configuration files serve as the team's "AI operations manual," allowing new members to reach the team's average level without starting from scratch.

Tips 9 & 10: SDK and Multi-Instance Concurrency

Using Claude Code SDK

If you want to further unleash Claude's potential, try the Claude Code SDK. It allows you to access Claude Code programmatically, use it in CLI environments, integrate it into your own applications, and even perform piped input/output with UNIX tools.

Claude Code SDK supports subprocess mode, enabling piped communication with other programs through standard input/output, fully compatible with the compositional tool design in UNIX philosophy. Developers can write scripts in Python, JavaScript, and other languages, embedding Claude Code as an "intelligent function" into CI/CD pipelines, code review tools, or custom development environments. SDK output supports structured JSON format for easy downstream parsing and processing.

For example: you can use the SDK to have Claude analyze Git status — it will explain status changes and output results in JSON format — opening up vast possibilities for automation workflows and secondary development.

Running Multiple Claude Instances Simultaneously

Boris listed four ways to run Claude concurrently. One team member even developed a small tool to monitor the working status of multiple Claude instances and multiple command lines, displayed as red/yellow/green traffic lights. Since screen space is limited, showing just the title bar is enough to grasp each process's running status at a glance.

Multi-instance concurrency faces unique challenges in AI coding agent scenarios: multiple instances may simultaneously modify the same file causing conflicts, or make contradictory architectural decisions without knowing each other's progress. Boris's team's practice is to strictly divide tasks along module boundaries, ensuring different instances operate on different file sets while isolating each instance's workspace through Git branches, with merging performed afterward. This pattern is similar to human team division of labor, but can execute several times faster.

The value of multi-instance concurrency lies in: allowing different Claude instances to simultaneously handle tasks for different modules, dramatically shortening the overall development cycle.

Summary: The Core of AI Coding Efficiency Lies in Methodology

Although Claude Code currently has certain usage restrictions for accounts in some regions, the workflow approach Boris shared remains highly instructive.

The core message it conveys is: AI coding efficiency doesn't depend on how powerful the tool itself is, but on how you set it up, how you ask questions, how you plan, and how you collaborate. From optimizing settings to team sharing, from planning first to multi-instance concurrency, these 10 tips are essentially a complete "human-AI collaboration methodology" applicable to any AI coding tool.

From a broader perspective, this methodology marks a fundamental shift in the software development paradigm: developers' core competitiveness is shifting from "the ability to write code" to "the ability to orchestrate AI work." Just as the Industrial Revolution transformed manual laborers into machine operators, the AI coding era is transforming code writers into directors and architects of AI programming. Developers who master this methodology will create orders-of-magnitude efficiency gaps compared to traditional approaches.

Whether you're using Claude Code, GitHub Copilot, Cursor, or other tools, switching your mindset from "writing code yourself" to "guiding AI to write code" is the true key to doubling your efficiency.