A practical guide to getting great results from Claude while keeping token usage — and cost — minimal.
Claude is a powerful AI assistant that can help you with drafting, coding, data analysis, and research.
To ensure our organization remains cost-effective and that everyone has continuous access to the tool, it is important to use Claude efficiently. Every time you send a message or upload a file, you consume tokens — the small chunks of text (roughly ¾ of a word, or about 4 characters, each) that Claude reads and writes.
This guide will help you get the best results from Claude while minimizing unnecessary token usage.
1 Choosing the Right Model
The #1 way to stay efficient — match the model to the job.
Your Default
Claude Haiku 4.5
Fast, cost-effective, and highly capable. Use Haiku for ~90% of daily tasks: drafting emails, summarizing documents, extracting information, formatting text, and basic brainstorming.
Available Upgrade
Claude Sonnet 4.6
Slower and more resource-intensive, but smarter. Switch to Sonnet 4.6 only when you need complex coding, deep logical reasoning, or analysis of large / highly complex datasets.
💡 Tip: If you switch to Sonnet for a tough problem, switch back to Haiku once the heavy lifting is done. And when in doubt, stay on Haiku — only escalate to Sonnet after Haiku has actually struggled with the task, not preemptively.
2 Smart File Management
Uploading massive or numerous files consumes a huge number of tokens instantly.
Don't upload "data dumps": Instead of uploading an entire 200-page PDF to ask one question, extract the relevant pages or chapters and upload only those.
Copy-paste text instead of screenshots: Uploading an image of a document consumes far more tokens than copying and pasting the raw text.
Prefer text / CSV over PDF, DOCX, or images: A native text or CSV file tokenizes much more cheaply than a scanned PDF or an image of the same content.
Consolidate files: If you have 5 small text files, combine their contents into one document before uploading. One file is more efficient than five.
Clean your data: Before uploading CSVs or Excel files, remove empty columns, irrelevant rows, and heavy formatting.
3 Mastering "Context" (Chat History)
Claude re-reads the whole conversation on every message — so long chats get expensive.
Start a new chat for new topics: Don't ask a coding question in the same chat where you drafted a marketing email. A fresh chat wipes the slate clean.
Avoid "Frankenstein" chats: If a single chat becomes very long (e.g. 30+ back-and-forth messages), it's time to start a new one.
Use the "Summarize & Restart" trick: In a long chat, ask Claude: "Please summarize our progress and the current state of this project in 3 paragraphs." Then copy that summary, start a New Chat, paste it in, and continue from there.
💡 Claude has built-in handling for long conversations, but a manual summary + new chat is still the most reliable way to control cost.
4 Prompting Efficiently
Clearer prompts mean fewer do-overs — and fewer wasted tokens.
Be direct and specific
✗ Vague"Help me with this email."
✓ Specific"Proofread this email for typos and make it sound more professional: [paste email]."
Control output length: Tell Claude exactly how much you want — "Summarize this in exactly 3 bullet points," or "Keep your response under 200 words."
Don't ask Claude to repeat things: When rewriting a paragraph, add "only output the revised paragraph, do not include the rest of the text." This saves output tokens.
Provide examples: If you want a specific format, show one example. This stops Claude from guessing wrong and wasting tokens on do-overs.
5 Use Projects & Project Knowledge
Stop re-uploading the same reference files into every chat.
Upload reference material once: If a set of documents is used repeatedly, add them to a Project instead of re-attaching them to each new chat.
Keep chats short: With shared context living in the Project, your individual chats stay lean — you avoid paying the upload and context cost over and over.
6 Turn Off "Extended Thinking" for Simple Tasks
Extended thinking makes Claude reason at length before answering.
Great for hard problems — multi-step logic, tricky debugging, complex analysis.
Wasteful for routine work — reformatting a list, quick rewrites, simple lookups. Leave it off for everyday tasks and only enable it when a problem genuinely needs deep reasoning.
7 Protect Sensitive & Regulated Data
Efficiency never comes before security.
🔒
Do not paste passwords, API keys, secrets, customer PII, or confidential financial data into Claude unless our organization's data-handling policy explicitly permits it. When in doubt, leave it out.
8 Reuse Proven Prompts
Getting it right on the first try is the cheapest path of all.
Build a shared prompt library: Keep a small collection of prompts that reliably produce good results, so nobody burns tokens rediscovering what already works.
Adapt, don't reinvent: Start from a known-good prompt and tweak it, rather than trial-and-error from scratch.
9 Quick Do's & Don'ts Cheat Sheet
Print it, pin it, share it.
✅ DO
❌ DON'T
Keep Claude Haiku 4.5 as your default model.
Use Sonnet 4.6 for simple tasks like drafting emails.
Start a new chat when you change topics.
Keep one massive chat running for weeks.
Copy-paste the specific text you need.
Upload whole books or 10 large files for one data point.
Tell Claude to keep responses brief.
Let Claude generate 5 pages when you need a summary.
Clean spreadsheets before uploading.
Upload screenshots of text you could paste instead.
