When building AI‑powered applications, the cost of tokens can quickly become a bottleneck. This article distills the key concepts from a recent video by Vladimir Karpukhin, focusing on how to keep Claude, Codex, DeepSeek Harness, GLM, and Kimi within their limits. We’ll cover token fundamentals, context‑window tricks, project‑level prompt hygiene, and sub‑agent strategies that save both time and money.
Context and practical value
The video explains token limits for Claude, Codex, DeepSeek Harness, GLM, and Kimi, covering input/output token costs, context‑rotation commands, prompt hygiene, system prompt design, sub‑agents, and subscription strategies.
This article translates the video’s practical advice into a structured guide, adds a clear step‑by‑step workflow, and provides a FAQ to address common uncertainties.
Tutorial profile
- Format: how_to
- Topic: token-efficiency
- Audience: intermediate
- Tools: Claude, Codex, DeepSeek Harness, GLM, Kimi
Key takeaways
- Understand the difference between input and output tokens and how language choice affects cost.
- Use context‑rotation techniques like /clear, /compact, and side‑chat to keep conversations lean.
- Clean project metadata (MCP, skills, CLAUDE.md, AGENTS.md) to reduce prompt noise.
- Apply Project Index, Search First, Frugal Reading, and Deny Noise principles to craft efficient system prompts.
- Leverage sub‑agents for research, coding, and review to distribute token usage.
Token Basics and Language Impact
Tokens are the smallest units of text that AI models process. The video explains how English typically consumes fewer tokens than Chinese for the same content, and how this affects cost. Knowing the token cost per model helps you choose the right language for your prompt.
Managing Context Windows
Long chats inflate token usage. The author introduces context‑rotation commands such as /clear and /compact, and suggests opening a new dialogue for fresh topics. Side‑chat usage keeps the main thread focused and reduces token waste.
Project Prompt Hygiene
Cleaning up MCP, skills, CLAUDE.md, and AGENTS.md removes unnecessary instructions. A tidy prompt reduces noise and keeps the model focused on the task.
System Prompt Design
The article outlines a framework: Project Index, Search First, Frugal Reading, and Deny Noise. These principles guide the creation of a lean system prompt that still delivers comprehensive guidance.
Sub‑Agents for Efficiency
Dividing work among research, coding, and review agents spreads token usage and speeds up iteration. The video demonstrates how to set up these roles for maximum savings.
Subscription and Harness Strategies
Choosing the right subscription tier and harness (Claude Code, Codex, Kimi Code) can affect token limits. The author discusses the five‑hour window of Claude Code and introduces Claude Routines to pre‑seed conversations.
Practical next steps
- Calculate token cost per model and set a budget before starting a project.
- Clean all project metadata files to remove redundant instructions.
- Implement context‑rotation commands in your workflow to keep conversations short.
- Apply the Project Index framework to your system prompt.
- Create sub‑agents for distinct tasks and route prompts accordingly.
Limits and verification
- The article relies on the author’s subjective experience; real‑world token usage may vary by model version and usage pattern.
FAQ
What is the difference between input and output tokens?
Input tokens are the text you send to the model; output tokens are the text the model returns. Both count toward your limit, but output tokens are often more expensive.
How can I measure token usage in my own projects?
Most APIs expose a token count in the response metadata. You can log this value for each request to track usage.