MODEL & SETTINGS
01
Use Sonnet, not Opus
Opus is the Ferrari. Sonnet is the ute that does everything you actually need. For emails, drafts, summaries, and brainstorming — Sonnet is your tool. It burns through your usage allocation far more slowly. Save Opus for the work that genuinely earns it.
02
Turn off extended thinking
Extended thinking is Claude doing its homework before answering. Useful when you're solving something hard. Complete overkill for "write me a follow-up email." It uses significantly more tokens, so leave it off by default and turn it on when you actually need the depth.
03
Disable connectors you're not using
Gmail, Drive, Calendar — every active connector adds context to every message, even when you're not using them. If it's not relevant to this session, turn it off. Takes three seconds. Saves a lot of unnecessary overhead.
HOW YOU PROMPT
04
Start a new chat for each new task
Long threads carry weight. Every message you send in an old conversation is dragging the whole history with it. New task, new chat. It's faster, cleaner, and doesn't cost you anything to start fresh.
05
Be specific from the first message
"Write me something about my business" is not a brief — it's a wish. Tell Claude the format, the audience, the length, the purpose. Right up front. Every round of back-and-forth you avoid is time and tokens back in your pocket.
06
Batch your asks into one message
Need three things done? Ask for all three in one message. Not three separate ones. One prompt, one response, done. Running three separate messages for one job is like making three trips to the kitchen.
"Rubbish in, rubbish out. The clearer your brief, the less you'll need to ask twice."
FILES & DOCUMENTS
07
Paste the text, don't upload the document
Uploading a full document means Claude loads the whole thing into context. When you only need two paragraphs from a 60-page report, copy and paste those two paragraphs. Claude doesn't need the rest — and processing it costs you.
08
Use Projects for files you come back to — but keep them lean
If you're working with the same files regularly — your rate card, your service list, your standard template — a Claude Project means you don't have to re-upload them every session. But here's the important bit: files in a Project still load into context when you work. Large files still consume memory on every task. Keep Project files tight — summaries over full documents where you can. If a file no longer earns its place, delete it.
09
Keep your project instructions short — unless the work demands it
Your project instructions load on every single task. Every unnecessary word is overhead you're paying for every time. For most workflows, under 500 words is the right target. The exception: complex setups or tool-heavy configurations may warrant more. The principle stays the same — trim ruthlessly. If a sentence doesn't change how Claude behaves, it doesn't belong there.
10
Remove files you no longer need
Old documents sitting in a Project can still be pulled into context even when they're not relevant. Do a clean-out every few weeks. Less stored in your Project means cleaner, faster, more focused responses — and less memory being used on things that don't help the current task.
WORKING SMARTER
11
Build a Master Prompt once
If you're re-explaining who you are at the start of every new session, stop. Write a Master Prompt — your role, your business, your tone, your standards — and paste it in at the top of any fresh chat. Two minutes of setup. Zero re-explaining. Claude goes from cold start to up to speed in seconds.
12
Build role prompts for tasks you do every week
Same kind of task every week? Build a role prompt for it once and keep it. A good role prompt tells Claude exactly who it's being, what the output looks like, and what the rules are. Paste it in, add your specifics, go. No rebuilding from scratch.
13
Use constraints, not just descriptions
Most people tell Claude what they want. Smarter move: also tell it what to leave out. "Under 150 words. No intro. No bullet points. Get straight to the point." Constraints get you there in one round. Descriptions take three.
