How To Never Run Out Of Codex and Claude Usage Limits

Summary

AI Labs presents a three-level guide to managing Claude Code and Codex usage limits. Its aim is not just to make a plan last longer, but to preserve answer quality while reducing unnecessary context and model work. The video says it covers 24 tips; the subtitles describe the following main practices.

How Context and Usage Limits Work

• A context window holds the information available to the model for a conversation. The model receives the conversation history again with each request, so later messages can use more tokens than earlier ones.
• Context can also include project instructions—CLAUDE.md in Claude Code or AGENTS.md in Codex—plus tool and skill descriptions.
• A usage limit is separate from the context window: it limits total use across sessions and models. The presenters describe five-hour and weekly limits, with the five-hour countdown starting at the first message. Switching models does not restore usage already spent.
• The presenters also explain that compacting a conversation requires the agent to read and summarize it, which itself uses tokens.

Level 1: Workflow Habits

1. Clear context between unrelated tasks. Start a fresh context rather than carrying irrelevant history into a new job.
2. Compact only when needed. Summaries can lose important details and cost usage to create. If compacting, specify what must be retained; the presenters recommend doing it before a long break rather than after.
3. Rewind mistakes instead of layering corrections. Remove the mistaken exchange where possible, then issue a corrected request. The video describes using Escape twice in both tools, with an additional rollback command in Claude Code.
4. Combine related work into one well-structured request. This can reduce repeated context and let the agent plan several changes together.
5. Turn off automatic session recaps if you don’t need them. In Claude Code, the recap uses a model request; the presenters suggest generating one manually when useful.

Level 2: Tool and Project Settings

1. Match the model to the task. Use smaller, faster models for routine work and reserve more capable models for harder tasks.
2. Keep effort or reasoning at a moderate default. Increase it for tasks that genuinely need deeper analysis rather than using a high setting all the time.
3. Route subtasks to suitable subagents or models. The presenters describe specifying a model in a request or using a “model router” skill.
4. Keep skills focused and load them only when relevant. The video recommends short skill files, with detailed material in separate reference files that can be loaded as needed.
5. Keep project instructions concise. Put only essential, project-specific information in the main instruction file; move specialized guidance into skills.
6. Use folder-specific instruction files on large projects. This can avoid loading rules for unrelated parts of a codebase.
7. Check usage regularly. A usage command is described as showing consumption in the five-hour and weekly windows, and identifying costly areas such as skills or tools.
8. Consider a CLI instead of an MCP when both are available. The presenters’ reasoning is that MCP tool descriptions remain in context, while a CLI is invoked through the terminal when needed. They use Supabase as an example and explain that skills can teach the agent how to use its CLI.

Level 3: Reduce Background and Tool-Output Costs

1. Review or disable memory if it adds irrelevant context. The presenters say stored memories can be included in future sessions.
2. Disable bundled skills you rarely use. This reduces the set of built-in skills available to the agent.
3. Disable Claude Code workflows, or limit their size, if unnecessary. The video says workflows can launch many subagents and use substantial usage.
4. Filter command output with a hook. Their example keeps failed test output while filtering successful results that the agent does not need to act on.
5. Set limits on tool output. The presenters describe reducing Claude Code’s maximum bash output length and setting Codex’s tool-output token limit.
6. Use a stronger-model advisor selectively. The video says Claude Code’s automatic advisor may resend the conversation to a stronger model; the suggested alternative is to call a consultant agent only when requested.
7. Coordinate named sessions with targeted messages. When sessions communicate, send only the information the other session needs rather than transferring all the context.

Guide and Examples

This is a tutorial and configuration guide, not a product review. It explains context and usage limits, recommends workflow and settings changes, and walks through examples involving project instruction files, skills, Supabase’s CLI/MCP options, test-output filtering hooks, and output limits.

Main speakers/source: The AI Labs channel’s presenters, speaking as members of a software development company; no individual speakers are identified in the subtitles.