Five Token-Saving Strategies Keep Claude Pro Users Off the Max Plan

By Sophie Lindqvist

A single /compact command run on an expired chat cache consumed 11% of one Claude Pro user's five-hour limit in a single stroke. That measurement, reported by practitioner Simon Scrapes, anchors a set of five strategies that let Pro subscribers stay within their plan instead of paying for the more expensive Max tier.

The core problem is mechanical. Claude's prompt cache expires after one hour of inactivity. Once it does, revisiting an old chat forces the model to reread the entire conversation at full cost — up to five times the token usage of a cached session. Every technique below attacks that cost structure from a different angle.

Why is /compact a trap?

Summarizing a long chat with /compact works well while the cache is active. After expiry or invalidation, the command becomes expensive. The user who lost 11% of a five-hour window discovered this directly.

The /clear command is free but erases all context, which makes it unsuitable for ongoing projects. Scrapes recommends a third path: save critical context into a file such as "handoff.md" before clearing the chat. Anthropic itself works this way, using files like "claude_progress.txt" or GitHub commits to preserve session continuity.

Commands such as /loop, task scheduling and /goal carry a similar penalty. Each turn resends the entire chat to the model, so limiting their use keeps token consumption in check.

What does cache discipline change?

Beyond the one-hour expiry, switching models or effort levels mid-task invalidates the cache and forces a full reread. Users should set the model and effort level at the start of a session and leave them alone. Running Max effort on simple tasks adds avoidable reasoning costs.

The recommended habit is saving key session details into files at the end of each conversation. A pricing discussion, for instance, can live in a "pricing.md" file that Claude references in a fresh session — relevant context preserved, irrelevant history not resent.

How much does pre-loaded context cost?

Every new chat begins with pre-loaded data: the "Claude.md" file, installed skills, plugins and memory files. Users pay for that data with every message. Scrapes offers four reduction steps:

• Keep "Claude.md" under 200 lines.
• Move recurring workflows, such as weekly reports, into skills instead of conversation threads.
• Uninstall unused plugins.
• Archive completed tasks and old project notes from "Claude.md" into separate files.

A recent Anthropic desktop update now displays pre-loaded context at the start of each chat, so users can see exactly what they are paying for. Multi-context processes (MCPs) have also stopped being a burden: tools now load only when Claude requires them.

What does /usage reveal?

Available to both Pro and Max subscribers, the /usage command reports weekly data, flags items consuming more than 10% of the limit, and proposes targeted changes — trimming "Claude.md", for example, or addressing requests that exceed 49,000 tokens. Pressing "F" inside the interface lets Claude apply the recommendations automatically.

The command also tracks cache misses, including those caused by tool definition changes, and breaks down token usage per iteration for loops and sub-agents in long-running tasks.

When do sub-agents pay off?

Sub-agents suit tasks that generate large outputs. Anthropic's own demonstration runs a test suite and reports only failing tests with their error messages. Because sub-agents share the same usage limit as the main agent, they burn tokens quickly on small jobs; the main agent alone is often more efficient, and users should let Claude decide when delegation is warranted.

One tuning lever matters for long tasks: by default, sub-agents retain their cache for only five minutes. Extending the sub-agent prompt cache TTL to one hour through environment variables prevents repeated uncaching and recaching during pauses, improving performance and cutting cost.

Taken together, command restraint, file-based context, lean pre-loaded data, /usage audits and disciplined sub-agent configuration proved sufficient for Scrapes to conclude that the Pro plan meets most needs — evidence that careful cache and context management, not a plan upgrade, is the cheaper lever for heavy Claude users.