TOKENS: cómo AHORRAR en Claude y usar mejor la IA 🚀

Summary

The episode explains how to use Claude more efficiently when usage is limited by rolling time windows or token budgets. Although the examples focus on Anthropic’s Claude apps—including Claude Chat, Cowork, and Code—the speakers say many of the ideas also apply to other AI platforms.

Tokens, context, and usage limits

• Tokens are pieces of information a model processes. Text, images, audio, video, uploaded files, and tool results can all contribute to token use.
• Claude’s context window includes more than the latest prompt. It can include the full conversation history, Claude’s previous replies, hidden system instructions, files, tool outputs, MCP connections, and enabled Skills.
• As a conversation grows, earlier content may be sent again with later messages. Long chats can therefore use more resources, and the speakers cite research suggesting that model performance may also decline as context grows.
• The episode describes separate five-hour and weekly usage limits. Users can view usage in Claude’s Settings under Usage.

Practical ways to reduce usage

Inspect context with /context

The /context command reports what is taking up space, including system prompts, tools, Skills, and messages. The speakers demonstrate it in Cowork and Claude Code, noting that the display differs between the two.

Use /compact or start fresh when appropriate

/compact summarizes a long conversation so work can continue with a smaller context. The guest notes that users may have limited control over what the automatic summary preserves, so important details can be restated.

In Claude Code’s CLI, /clear clears the context for a new task.

Disable Skills and integrations you do not need

Skills, connectors, and plugins can load into context even before a user sends a message. In one demonstration, enabling around seven Skills increased reported context use from roughly 1,800 to 3,300 tokens. The speakers advise reviewing settings and checking whether plugins also enable Skills.

Choose a model suited to the task

The speakers’ general guidance is:

• Haiku: Fast, lower-cost tasks such as translation, summaries, or drafting a simple email.
• Sonnet: A good default for many work tasks, including document analysis and tool use.
• Opus: More complex planning or reasoning tasks.
• “Fable”: A still more expensive model referred to by this name in the subtitles, which may be mistranscribed. The guest says it is more costly and may require a separate budget on some plans.

The episode gives approximate relative costs of 3× Haiku for Sonnet, 5× for Opus, and 10× for “Fable.” These are the speakers’ comparisons, not universal prices.

One suggested workflow is to use a stronger model to plan a complex task, then hand the defined steps to a cheaper model to execute. Claude Code’s /plan mode is mentioned for planning before execution.

Use prompt caching and work in sustained sessions

The speakers describe caching as reusing previously processed content rather than processing it from scratch. They say Anthropic manages prompt caching automatically in its apps.

The guest cites an apparent cache lifetime of about five minutes and estimates possible savings of 40–80%, depending on the situation. These figures are estimates from testing and research discussed in the episode. Long pauses may mean cached content must be processed again.

Use Projects for recurring work

Projects can provide reusable files, instructions, and project-specific memory across chats, reducing the need to re-explain a task or re-upload reference materials. The speakers distinguish:

• Local projects, linked to a computer folder and limited by local storage.
• Online projects, which can be shared but have upload limits.

They describe interface differences between Chat and Cowork for local versus online projects, while cautioning that behavior may change.

Convert reference documents to Markdown where suitable

Plain-text formats such as Markdown can use fewer tokens than formats that include layout and other file information. In the episode’s example, a table used about 736 tokens as a PDF and 156 as Markdown—roughly five times fewer.

The speakers advise against converting sensitive or private company information using external services.

Batch questions and automate repeated work

Instead of sending several separate messages that each carry the growing conversation history, combine related questions into one well-structured prompt. For repetitive tasks, scripts or available tools can avoid repeatedly asking the model to recreate the same code or process.

Other tools and claims discussed

For Claude Code, the guest mentions community tools intended to reduce output or code volume:

• Pony Tails: Described as helping generate more efficient code. The episode cites claimed reductions in generated code of 54–94% in tested cases.
• Caveman: Encourages shorter, more direct responses. The episode cites about 33% fewer tokens.
• Graphify: Represents a codebase as a graph to help the model navigate its structure. The episode cites a claimed 71% reduction in token costs.

These are claims reported by the guest from the tools’ materials or tests, not independently verified results in the episode.

The speakers also discuss reports that Claude usage may count more heavily during peak hours. They cite a social media post by an Anthropic employee and mention time ranges in Pacific and GMT, but say they could not find an official, quantified rate.

Guides and demonstrations featured

• Walkthrough of the Usage settings and /context command.
• Demonstrations of enabling and disabling Skills and observing the context change.
• Guidance on using /compact, /clear, and Claude Code’s /plan.
• Comparison of Haiku, Sonnet, and Opus for different task types.
• Demonstration comparing PDF and Markdown token use.
• Examples of setting up Projects for recurring work.

Main speakers and sources

• Host: EvoAcademy presenter; name not stated in the subtitles.
• Guest: Jonathan, discussing Claude usage and token-saving workflows.
• Additional sources mentioned: Research on context-window performance, an Anthropic employee’s post about peak-hour usage, and the community tools’ GitHub pages or reported evaluations.