Token Conservation &
Context Management

Why Tokens Matter

Every time you interact with Claude, you're spending tokens — the unit of text that AI models use to read and write. Every word you send and every word Claude writes back costs tokens.

Think of tokens like words, roughly. One token is about three-quarters of an English word. The sentence "What is the capital of France?" is about 8 tokens.

Token conservation matters for three reasons:

How Token Costs Actually Work

Before we dive into tips, let's demystify what's happening behind the scenes. Most people think of Claude conversations like texting — you send a message, you get a reply. The reality is very different.

What Gets Sent Every Time You Hit Enter

Every single message you send includes everything — not just your latest text, but the entire conversation replayed from the beginning:

The key shock: by message 15 in a conversation, the chat history alone can be 60-90% of your token cost. Your actual new message might be only 2-10% of what gets processed.

What Each Plan Actually Gets You

Anthropic doesn't publish exact token budgets — they use a rolling window system. But based on community tracking, here are practical estimates per 5-hour window:

Estimates based on short-medium conversations. Long conversations, file uploads, and extended thinking reduce these significantly.

Quick Rules of Thumb

Prompt Caching: The Automatic Savings

Here's the good news: Claude automatically caches the parts of your conversation that don't change. On your second message, the system prompt and first exchange are cached — Claude reads them at 90% off instead of reprocessing them from scratch.

Caching happens automatically on claude.ai and in the API. You don't need to do anything to benefit from it. But it only helps with the repeated parts — your new messages and Claude's new outputs still cost full price. This is why shorter, more focused conversations are still cheaper even with caching.

Extended Thinking: The Hidden Cost Multiplier

When Extended Thinking is enabled, Claude "thinks" before responding. This thinking is billed as output tokens — the expensive kind (5x input cost). A single response with deep thinking can burn 5,000-50,000 thinking tokens on top of the visible response.

Conserve Your Input Tokens (What You Send)

Input tokens are what you send to Claude — your prompt, any pasted text, uploaded files, and your conversation history. Here's how to be efficient:

Before & After: Watch the Difference

Edit Your Prompt — Don't Send a Follow-Up

When Claude doesn't get it right, resist the urge to type "No, I meant..." as a new message. Every follow-up gets added to the conversation history, and Claude re-reads all of it on the next turn.

Instead: click Edit on your original message, fix it, and regenerate. The old exchange gets replaced, not stacked. This alone can save massive amounts of tokens over a session.

Batch Your Questions

Three separate prompts = three full context reloads. One prompt with three tasks = one context reload. You save tokens twice: fewer reloads, and you stay further from hitting your limit.

Bonus: batched answers are often better because Claude sees the full picture immediately.

Control Your Output Tokens (What Claude Writes)

Output tokens — what Claude writes back — are typically more expensive than input tokens. And Claude tends to be thorough by default. You can (and should) steer this.

1. Tell Claude the length you need. Say "in 2-3 sentences" or "keep it under 100 words." Without guidance, Claude will write as much as it thinks is helpful — which is often more than you need.
2. Specify the format. "Give me a table," "respond in bullet points," or "just give me the answer, no explanation." Format control is length control.
3. Ask for the answer first. If you say "Explain X," Claude will explain in detail. If you say "What is X? One line. Then explain briefly if needed," you get a concise answer with optional depth.
4. Use "concise" or "brief" as modifiers. Simply adding "Be concise" to your prompt can cut output length by 40-60% with no loss of usefulness.
5. Don't ask Claude to "explain your reasoning" unless you need it. This phrase alone can add hundreds of tokens to every response.

Turn Off Features You're Not Using

Web search, connectors, and "Extended Thinking" all consume tokens even when you don't need them. If you're writing your own content, turn off search. If your task is straightforward, keep extended thinking off — only enable it when your first attempt wasn't satisfactory.

Choose the Right Model

Not every task needs the most powerful model. Using the right model for the right task is the single biggest lever for managing your usage limits.

Extended Thinking (sometimes called "Ultrathink") follows the same principle. It's powerful for complex reasoning and architecture decisions, but leaving it on for routine tasks multiplies your token usage for zero benefit.

Master the Context Window

Claude's context window is its working memory — the total amount of text it can see at once. Think of it as a desk: the more papers you pile on, the harder it is to find what matters.

The Hidden Math of Long Conversations

Token cost isn't linear — it's quadratic. Each new message costs more than the last because Claude re-reads everything that came before. Here's what that looks like at ~500 tokens per exchange:

One developer tracked their usage and found that 98.5% of tokens were spent re-reading conversation history. Only 1.5% went toward generating the actual response. This is why starting fresh is so powerful.

Signs Your Context Is Getting Full

• Claude "forgets" instructions you gave earlier in the conversation
• Responses become generic or repeat things already discussed
• Claude contradicts something it said earlier
• Responses take noticeably longer and feel slower
• You get a warning that the conversation is approaching its limit

When to Start a Fresh Conversation

The Context Handoff Technique

When you need to start a new conversation but want to carry forward what you've done, use the handoff technique. This is the single most valuable skill for managing long-running work with Claude.

How It Works

1. Before closing the old conversation, ask Claude: "Summarize everything we've decided, produced, and any open questions — in a format I can paste into a new conversation."
2. Copy that summary. It should contain: decisions made, current state of the work, your requirements, and what's left to do.
3. Paste it at the start of your new conversation with a line like: "Here's context from our previous session. Continue from where we left off."

Use Projects, Memory & Preferences

If you find yourself typing the same instructions at the start of every conversation — your role, your writing style, how you want output formatted — you're wasting tokens. Claude has three built-in features to eliminate this repetition.

Set Up Memory & User Preferences

Every new chat without saved context wastes 3-5 messages on setup: "I'm a marketer, I write in casual style, I prefer short paragraphs..."

Go to Settings → Memory and User Settings. Save your role, communication style, and preferences once. Claude will automatically apply them to every new chat — zero tokens spent on repeat setup.

Upload Recurring Files to Projects

If you upload the same PDF, style guide, or brief to multiple conversations, Claude re-tokenizes it every time. Upload it once to a Project — every new conversation references it without burning extra tokens. For contracts, brand guides, or any long docs, this alone can cut your token spend dramatically.

Manage Your Usage Limits

Claude Pro and Team plans use a rolling usage window, not a daily reset. Understanding how it works gives you more effective capacity.

The Rolling Window

Claude uses a rolling 5-hour window. It doesn't reset at midnight — your usage gradually decreases as older messages "expire." Messages sent at 9 AM will no longer count toward your limit by 2 PM.

If you use up your entire limit in one intense morning session, you've left most of your daily capacity on the table. Instead:

1. Split your work into 2-3 sessions. Morning, afternoon, evening. By the time you return, your previous usage has rolled off and you have fresh capacity.
2. Front-load complex tasks. Start your day with the hard work that needs Opus or extended thinking. Use lighter models for afternoon tasks.
3. Avoid peak hours when possible. Usage during peak times (weekday mornings US time) may consume your limits faster. Running intensive tasks in the evening or on weekends stretches your plan further.

Built-In Tools You Should Be Using

Both claude.ai and Claude Code have built-in features specifically designed for token and context management. Most people never discover them.

Quick Reference Cheat Sheet

Bookmark this section. These are the highest-impact habits for everyday Claude usage.

Advanced Patterns for Power Users

The Progressive Disclosure Pattern

For complex, multi-step projects, don't dump everything on Claude at once. Feed context progressively:

The Checkpoint Pattern

For long workflows, ask Claude to create periodic checkpoints:

The Scope Limiter

When you don't want Claude to over-elaborate, set explicit boundaries: