How to Never hit Claudes Usage Limit (UPDATED!)

Summary

The video explains how to reduce Claude usage by managing tokens and context, rather than focusing only on the number of messages sent. Claude Code repeatedly processes conversation history and loaded project instructions, so long sessions and oversized setups can make each new prompt more expensive.

How usage accumulates

• Tokens are the underlying unit of usage. The speaker estimates that a token averages about four characters, or roughly three-quarters of an English word.
• Claude reprocesses the session’s instructions, conversation history, files, and tools with each new message. Usage can therefore grow rapidly in long chats.
• The context window is the temporary workspace containing this material. A large context can also contribute to context rot: the speaker says performance may worsen as important details become harder to find.
• The video recommends checking context usage in Claude Code and starting a fresh session around 120,000 tokens, rather than waiting for automatic compaction. It cautions that automatic summaries may omit useful technical details.

Reduce the context loaded into each session

• Disable MCP servers that are not needed for the current task; their tool definitions create recurring context overhead.
• Where practical, use CLI commands for occasional tasks, such as fetching web or API data, instead of keeping an MCP server active.
• Keep claude.md concise by removing duplicate, contradictory, obsolete, vague, or unnecessary instructions.
• Use progressive disclosure for long references: move detailed material into separate files and leave a short pointer in claude.md.
• Remove unused skills and trim unnecessary comments or explanatory text from those that remain.
• Adjust settings.json to compact earlier, allow longer command output to avoid repeated log requests, and exclude large, low-value directories from scanning. The exact configuration snippets are not legible in the subtitles.

Improve prompting and session workflows

• Edit the original prompt instead of sending follow-up corrections. This avoids keeping failed answers and correction exchanges in the active conversation.
• When a chat becomes heavy, ask Claude for a summary of the work and modified files, then continue in a fresh session.
• For larger projects, separate work into discovery, planning, and execution sessions. Carry forward a concise brief or implementation manifest rather than the full research history.
• Batch related requests into one prompt when possible, reducing repeated processing of the same context.
• Use Plan mode for complex tasks so the approach can be reviewed before code is generated.
• Use /rewind to return to an earlier point in a Claude Code conversation and discard a failed branch. The video also describes summarizing useful context from that point.
• Delegate large log or document analysis to a sub-agent, preferably using a cheaper model for summarization, and bring only the result back to the main session.

Choose models and files strategically

• Use Haiku for simple checks, formatting, and summarization; Sonnet for most routine coding; and Opus for more demanding architectural or complex-logic work.
• Convert bulky reference files to Markdown where possible. The speaker says this can reduce token overhead compared with HTML, PDF, or DOCX while retaining useful content, and demonstrates using Pandoc.

Timing, monitoring, and additional resources

• The video describes Claude usage as operating on a five-hour rolling window, with capacity returning gradually as earlier usage ages out. It recommends taking breaks and spacing out intensive work.
• The speaker also claims that off-peak periods may be more favorable, though the timing and efficiency claims are not demonstrated in the subtitles.
• Pro and Max users are shown how to enable extra usage and set a spending cap as a backup.
• The speaker recommends a local dashboard that reads Claude Code logs to chart token use, sessions, and model usage, as well as a “context audit” tool that checks project instructions, files, skills, settings, and MCP servers.
• Other mentioned resources include GitHub collections of token-reduction wrappers, the Caveman plugin for concise responses, and the B-mat optimization framework.

Guides and tutorials covered

1. Check context usage in Claude Code and reset or hand off a session before it becomes too large.
2. Disable unused MCP servers and substitute CLI commands for occasional external lookups.
3. Audit and streamline claude.md, skills, project settings, and scanned directories.
4. Edit prompts, batch tasks, and divide project work into discovery, planning, and execution.
5. Use Plan mode, /rewind, and sub-agents to limit costly failed attempts and heavy analysis in the main session.
6. Convert reference documents to Markdown and monitor usage with a local dashboard or context audit.

Main speaker and sources

• Main speaker: The video’s narrator from the Mikey Vibe Coding channel.
• Sources and tools discussed: Claude and Claude Code, including Anthropic’s automatic compaction and usage limits; MCP servers; CLI tools; Pandoc; GitHub-hosted token-tracking and context-audit resources; Caveman; and B-mat.