1. Convert files before you upload them
A single PDF page costs 1,500 to 3,000 tokens. A full-screen screenshot is about 1,300. Upload the same 15-page PDF to four chats and you've burned 180,000+ tokens on something you could have flattened to 2,000 tokens of clean text.
- Extract the text first, then paste the parts that matter into a plain .txt or .md file.
- Crop screenshots tight. A tight crop drops from ~1,300 tokens to under 100.
Fast workflow: open a blank Google Doc (type doc.new in the address bar), paste your text, download as Markdown (.md).
2. Plan in Chat. Build the file at the very end
File creation (spreadsheets, docs, decks) eats far more of your limit than regular chat. So do the thinking in the cheap product and the building in the expensive one: plan the structure in Chat, lock the sections and assumptions, then move to Cowork and say “build this exact file.” You pay building prices once instead of ten times.
3. Match the model to the task
Opus + Extended Thinking is heavy machinery. Don't use it to move a chair.
- Quick question, grammar check, brainstorm? Haiku or Sonnet, at a fraction of the cost.
- Report from your files? Cowork with Opus.
Rule of thumb: if a task takes Claude under 30 seconds, it doesn't need Opus. Switching models is two clicks before you start.
4. Edit instead of piling on follow-ups
In Chat, hit Edit on your original message, fix it, and regenerate, so the old exchange is replaced, not stacked. Every “no, I meant…” or “actually, change X to Y” adds to history Claude re-reads forever. And when one section is wrong, say “only redo section 3, keep the rest.” A full redo re-generates every output token from scratch.
5. Batch your tasks into one message
Three prompts = three full context reloads. One prompt with three tasks = one reload. Instead of “summarize this,” then “list the main points,” then “suggest a headline,” send all three at once. Bonus: the answer is usually better, because Claude sees the whole picture, same as a human would want.
6. Summarize and start fresh every 15 to 20 messages
Long conversations are token furnaces. One developer measured 98.5% of tokens going to re-reading history and only 1.5% to actual output. When a session drags: ask Claude to summarize the essentials, copy it, open a new chat, paste it as message one. For scale: a 20-message session burns about 105,000 tokens; a 30-message session burns 232,000.
7. Use Projects for recurring files
Upload the same PDF to five chats and Claude re-tokenizes it five times. Use a Project instead: upload once, it gets cached, and every conversation references it without re-reading. On paid plans, Projects also use retrieval (RAG), so Claude pulls only the relevant chunks instead of loading the whole document. If you lean on contracts, brand guides, or research papers, this alone cuts spend hard.
8. Zoom out: find the waste across your whole stack with ELI
Here's the uncomfortable truth: Claude is one line item. You just learned to stop wasting tokens on one tool. Now look at everything else your company runs on: the ghost licenses, the duplicate tools, the shadow AI subscriptions nobody admits to. On average that's around half of a company's software budget, gone.
ELI is the control layer for every tool, model, agent, and subscription your company runs on. It connects to your inbox, accounting, and browser, then diagnoses the waste, cuts it, and governs what's left:
- Stack visibility: every app and account in one place, including the ones you forgot you pay for.
- Cost optimisation: ghost seats surfaced, renewals flagged weeks ahead, duplicate tools (Jira vs Linear, Mixpanel vs Amplitude) scored by ROI so you cut the biggest dollars first.
- An agent fleet: a dozen small agents watching onboarding, renewals, spend, off-boarding, and compliance, pinging you only when a human call is needed.
- An MCP gateway: drop your workspace endpoint into Claude or Cursor and your agents can query your real stack data, read-only and safe.
The pitch in one line: ELI is priced as a small percentage of what you already spend, and you pay only if it saves you money. Free scan, live in 90 seconds. Find out exactly where your stack is leaking.
Where to start
Don't try to do all eight at once. Pick the three that match how your team works:
- Live in Cowork? Habits 1, 2, and 4: convert files first, plan in Chat before building, edit instead of redoing.
- Mostly in Chat? Habits 5, 6, and 7: batch prompts, start fresh often, Projects for recurring files.
- Want the real win? Habit 8: run a free ELI scan and see what the rest of your stack is costing you.
Save this. Forward it to your team. Every habit you build is budget you keep for the work that actually matters.