18 Claude Code Token Hacks in 18 Minutes

Summary

This video dives into 18 practical hacks to drastically optimize Claude code token usage, helping users get more done without hitting session limits too quickly. Starting with beginner tips like clearing conversations, batching prompts, and monitoring token usage, it progresses to intermediate strategies such as maintaining a lean claud.md, compacting context, and limiting command output. Advanced users learn to choose the right models, manage sub-agents, and strategically schedule heavy sessions during off-peak hours. Throughout, the emphasis is on understanding how tokens work, maintaining context hygiene, and balancing quality with efficiency, empowering users to become high-performing power users of Claude code.

Takeaways

• 😀 Tokens are charged every time Claude reads your conversation, and costs compound as the session grows.
• 😀 Most token usage comes from rereading old chat history rather than generating new content.
• 😀 Starting fresh conversations with `/clear` prevents unnecessary token accumulation across unrelated tasks.
• 😀 Disconnecting unused MCP servers reduces hidden token overhead significantly.
• 😀 Batching multi-step instructions into one prompt saves tokens compared to sending multiple messages.
• 😀 Plan Mode helps prevent wasted tokens by letting Claude map out tasks before execution.
• 😀 Monitoring with `/context`, `/cost`, and a status line provides visibility into token usage and session health.
• 😀 Lean and precise files (like `claw.md`) and specific file references prevent Claude from processing unnecessary data.
• 😀 Sub-agents and high-cost models should be used sparingly and strategically, ideally during off-peak hours.
• 😀 Compacting context at around 60% and managing session breaks reduces token waste and maintains output quality.
• 😀 Watching Claude work and applying system-level rules in `claw.md` ensures efficient use of tokens and reduces repetitive work.
• 😀 Hitting your token limit can be a sign of productive usage if tokens are managed wisely; balance cost and quality.

Q & A

What is the primary reason people are hitting their Claude code token limit faster?

The primary reason is that Claude rereads the entire conversation history with each message, causing token usage to compound exponentially. This makes even small conversations much more costly as the session progresses.

How does token usage increase over the course of a conversation with Claude?

Token usage grows exponentially because Claude rereads every message in the conversation. For instance, the first message might cost 500 tokens, but by the 30th message, the cumulative token cost can reach 15,500 tokens, much higher than a linear increase.

What is the 'loss in the middle' phenomenon?

The 'loss in the middle' refers to a situation where Claude pays more attention to the beginning and end of a conversation, while ignoring much of the middle portion. This reduces the quality of responses for older messages in long sessions.

Why is it important to start fresh conversations and use the /clear command?

Starting fresh conversations with the /clear command prevents unnecessary tokens from being spent on rereading unrelated context. This is a simple but effective way to optimize token usage.

What are MCP servers, and how do they impact token usage?

MCP servers load additional context into your conversation on every message, which consumes tokens. Disconnecting unused MCP servers helps prevent this invisible overhead and reduces token consumption.

How can batching prompts into one message save tokens?

Instead of sending multiple messages for separate tasks, batching prompts into a single message reduces the overall token cost. Each separate message adds extra token usage because Claude has to process each one individually.

What is the role of plan mode in managing token usage?

Plan mode allows Claude to map out the task, ask necessary questions, and gain clarity before diving into the work. This prevents token waste caused by going down the wrong path or redoing work due to missteps.

How can the /context and /cost commands help in managing token usage?

The /context command shows what is consuming tokens in real time, such as conversation history and MCP overhead. The /cost command displays your current token usage and estimated spend, helping you track and control your token consumption.

What is the advantage of keeping your cloud.md file lean?

A lean cloud.md file ensures that Claude only loads essential context, avoiding unnecessary tokens spent on reading lengthy files. It should contain key rules and decisions, and point to external files as needed to keep token usage efficient.

Why should you be specific when referring to files in Claude?

Being specific with file references, like pointing to a particular function in a file, prevents Claude from wasting tokens by reading unnecessary parts of large files. This helps keep token usage efficient and focused on what’s needed.

What is the impact of sub agents on token usage, and how should they be used wisely?

Sub agents consume significantly more tokens than single-agent sessions because they load a separate full context. They should be used sparingly, preferably for specific tasks like research or summarizing large datasets, to save tokens while still benefiting from their capabilities.

How can understanding peak hours affect your token usage?

During peak hours (8 AM - 2 PM EST), token consumption can be faster due to higher demand. By scheduling larger tasks during off-peak hours, you can extend your session's lifespan and reduce token usage, maximizing the value of your plan.

What is the balance between quality and cost in token usage?

There is a trade-off between the quality of output and the cost in tokens. Sometimes higher quality tasks will require more tokens, but understanding when to prioritize quality or efficiency can help optimize token usage without unnecessary costs.

How can compacting context at 60% capacity help manage token usage?

Running the /compact command when your context reaches 60% capacity helps reduce bloat and manage token usage more efficiently. Compacting helps to preserve important context while discarding unnecessary data, keeping the session optimal.