A guide by Anatoli Kopadze that went around X (6.7k likes). The premise: most daily Claude users run at 10% of what it can do, not because it is hard, but because nobody showed them the rest. The fix is mostly setup and a few prompting habits, not clever tricks. Worth keeping as a checklist to hand to people who only use Claude as a search box.
Set it up once
- Use a Project, not a chat. A Project is a persistent workspace that keeps context across every conversation inside it. Set it up once and every session starts with Claude already knowing who you are.
- Tell Claude who you are. Drop a filled-in profile (work, goals, how you like to communicate) into the Project knowledge base. Specificity drives quality on every later response.
- Convert that into Custom Instructions. The profile says who you are; Custom Instructions say how Claude should behave with you by default. Have Claude generate them from your profile, then paste into Project Instructions.
Reframe how you use it
- It is a thinking partner, not a search engine. Treating it like Google cuts usefulness by ~80%. Stop asking what something is; ask it to help you think through a concrete problem. "Walk me through how prompt caching works given that I call Claude 20 times per session" beats "what is prompt caching".
- Ask Claude to ask you questions first. Before any complex task, tell it to gather requirements from you. The author rates this the single highest-leverage habit for fixing bad outputs.
- Style cloning. Give it 3-5 samples of your own writing and ask it to analyse the patterns, not just describe the style. Then it writes like you instead of a corporate assistant.
- Use it as a sparring partner. For any plan or decision, tell it to attack the idea, not critique it. Default behaviour is agreement and elaboration, which does not stress-test anything.
- Extended Thinking. The brain icon turns on step-by-step reasoning before the answer. Skip it for simple tasks, use it for hard decisions and analysis.
- Let Claude write prompts for Claude. If you are unsure how to prompt for a task, ask it to write the prompt. It knows what instructions produce better results.
Spend fewer tokens
- Specify output length up front. The default is longer than you need. Stating the length cuts token use 40-60% on most tasks with no loss of value.
- Kill the preamble. Add a Custom Instruction to drop "Great question", restatements, disclaimers, and closing summaries.
- Stop re-explaining yourself. Background goes in the Project once, read automatically every session. Never paste it again.
- New topic, new chat. Long chats keep all prior context loaded, which costs tokens and bleeds old context into the new topic. Start a fresh chat inside the Project to keep memory but drop the baggage.
Ready-to-use prompt patterns
The guide ends with copy-paste prompts for: explaining anything via analogies (Feynman method), travel planning around your actual style, monthly expense analysis from real data, structured self-reflection, and ruthlessly stress-testing a business idea before committing.
The closing line: Claude is not smarter than you, but it has infinite patience, broad knowledge, and angles you have not considered. The people who get the most from it are not the ones with the best questions, they are the ones who set it up to understand them.