How to Use Claude Code Effectively and Save Tokens

TL;DR

Use plan mode for large tasks, document project-specific guidance in CLAUDE.md, and create focused skills for repeatable workflows. To reduce usage and improve results, let a stronger model research and plan the work, then assign implementation to a faster, cheaper model, while avoiding unnecessary planning and excessive skill installation.

Transcript

I have spent an unhealthy amount of time using Claude code. And I got to get this out of my chest already. Their usage limits suck. Oh my goodness. I say hi and boom, there goes my limit. And I got to wait 3 hours. And with this new Claude Fable model. [laughter] Oh I recorded a lot of this footage like a month ago. So yes, I know it's not new. Lea... Read More

Key Insights

• Claude Code is a coding-focused version of Claude that primarily operates in a terminal, with additional options for IDE and web use. The creator prefers an IDE-and-agent combination because it provides both an agent interface and visual controls that can be clicked.
• Usage limits are a major constraint on the $20 plan because the creator reports sometimes reaching the limit after only one or two prompts and then waiting three hours. This limited interaction time can prevent users from learning how to operate the tool effectively.
• The slash init command is designed for the beginning of a project and creates a CLAUDE.md file after examining the codebase. That file acts as persistent project memory and can also import additional files, including an existing AGENTS.md file.
• A useful CLAUDE.md file contains concise project-specific context, including the project description, current status, coding style, working philosophy, capitalization preferences, and pull request language. Its contents should remain flexible because the file is helpful but is not essential for successful work.
• A skill is a SKILL.md guide that captures a repeatable workflow, set of best practices, or research process for an agent to apply when appropriate. Skills can also teach developers by exposing the knowledge and working methods contributed by companies and experienced engineers.
• Skill selection works best when it is restrained and consistent. Installing a large collection can introduce conflicting opinions about good code, so the creator recommends choosing one coherent group of skills or selecting skills that closely match the developer's own coding style.
• Plan mode works by making Claude inspect the code more thoroughly, prepare a complete plan, and request approval before changing anything. It is recommended for large tasks because reviewing a plan is easier than finding a mistake across thousands of generated lines and many modified files.
• A two-model workflow can reduce cost by assigning research and planning to a stronger model, then giving implementation to a faster and cheaper model. The creator suggests Opus for planning and Sonnet for implementation within Claude Code, while other coding agents can support broader model mixing.

Questions & Answers

Q: What is Claude Code and where can it be used?

Claude Code is described as Claude specialized for programming. It primarily lives in the terminal, but it can also be used inside an IDE and on the web. The creator prefers an IDE-and-agent combination because it supplies both an agent interface and a conventional development environment, including visual buttons, instead of requiring every interaction to happen through the terminal.

Q: How can developers reduce token usage in Claude Code?

Developers can conserve usage by reserving plan mode for substantial work, avoiding it for trivial edits, and splitting responsibilities between models. A stronger model can examine the codebase and write the plan, while a faster and cheaper model implements that plan. Focused project instructions and carefully selected skills can also reduce repeated explanations and unnecessary agent activity.

Q: What does the slash init command do in Claude Code?

The slash init command is intended to be run when Claude Code first opens a project. It examines the codebase and writes its findings into a CLAUDE.md file, which becomes persistent project memory. An experimental version changes the process by interviewing the user and recommending skills and hooks instead of relying only on automatic codebase analysis.

Q: What should a CLAUDE.md file contain?

A CLAUDE.md file can contain a short project description, the project's current status, preferred coding styles, and a working philosophy. The creator also uses rules for user-facing capitalization and specifies the language for pull requests. The file should evolve with experience, but it does not need to become exhaustive because a project can function without one.

Q: What are skills in Claude Code?

Skills are markdown guides stored specifically as SKILL.md files. They describe repeatable practices, workflows, or research methods that the agent can apply when relevant or when explicitly requested. Claude Code includes some default skills, such as code review and security review, while community skills provide additional expertise and can be educational for developers who read their instructions.

Q: How many Claude Code skills should a developer install?

A developer should avoid installing or invoking a very large number of skills. The creator warns that coding guidance can be subjective, and different skills may embody incompatible ideas about good code. A more consistent approach is to use one related family of skills or choose a small set that aligns with the developer's preferred coding style and workflow.

Q: When should plan mode be used in Claude Code?

Plan mode should be used for large or complex tasks that benefit from deeper codebase exploration before implementation. It allows Claude to prepare a complete plan for approval before making changes, which makes mistakes easier to identify. It should generally be skipped for typos, variable renaming, and small design adjustments because those tasks gain little from additional planning.

Q: How does the two-model planning workflow work?

The two-model workflow assigns planning to the more capable model and implementation to a faster, cheaper model. The first model audits or explores the codebase and produces detailed instructions for the second agent. Within Claude Code, the creator proposes Opus for planning and Sonnet for implementation. Other coding agents can offer additional flexibility by allowing models to be mixed.

Summary & Key Takeaways

• Claude Code is a coding-focused version of Claude that primarily lives in the terminal, although it can also be used through an IDE or the web. The creator finds it useful and relevant to employment, but criticizes the restrictive usage limits experienced on the $20 plan and prefers an IDE-plus-agent interface.

• The slash init command examines a project and creates a CLAUDE.md file that serves as persistent project context. Useful contents include a project description, current status, coding conventions, working philosophy, capitalization preferences, and pull request language. The file is helpful but optional because well-designed skills can provide comparable or better guidance.

• Skills store repeatable workflows and best practices in SKILL.md files, while plan mode makes Claude inspect a codebase and propose an approach before editing it. The recommended workflow uses planning for substantial tasks, reserves direct action for small changes, and combines a stronger planning model with a faster implementation model to conserve resources.