Try this right now. Open Claude and type: "Write me an email." You'll get a generic template with "I hope this message finds you well" and placeholder brackets everywhere.
Now try: "Write a follow-up email to a client named Sarah who missed our Tuesday call. Tone: warm but direct. Mention rescheduling for Thursday at 2pm. Keep it under 100 words."
Night and day. The difference is context. Claude can't read your mind, but it's exceptional at working with whatever you give it. Most people treat prompts like texting a friend who knows the backstory. Claude knows nothing until you tell it.
Claude needs five things from you:
1. Who (the audience). "A message to a VP you've never met" drives tone, vocabulary, and length.
2. What (the deliverable). Email, memo, analysis, script. Naming it lets Claude follow the right conventions.
3. How (tone, length, format). "Professional but warm, under 150 words, bullet points."
4. Why (the purpose). "Convince a skeptical CFO" produces different content than "brief a room that already agrees."
5. What not (exclusions). "Don't mention the price increase yet. Avoid corporate buzzwords."
A simple task needs one or two. A complex deliverable benefits from all five. The more you provide upfront, the less you edit after.
When do you start a new conversation? Continue when the new request builds on what came before. Start fresh when you switch topic, audience, or goal.
Claude builds on everything earlier in the conversation. Established tone, audience, and constraints stay active for every new message. That's powerful while you iterate on one deliverable, and a problem when you switch tasks, because old assumptions carry over.
Continue when: you're editing the same document, iterating on a draft, or asking follow-ups on the same topic.
Start new when: you're moving to a different task or want fresh context. The habit: one conversation per project.
The first response is version 0.5, not the final draft. Tell Claude what to adjust: "Make the opening more personal." "Too formal, strip the corporate language." Each round gets closer, and Claude remembers all feedback in the conversation.
Three patterns that consistently work:
The Layered Brief: full context in one message (role, audience, deliverable, constraints), then iterate on what's off. Use it when you know the requirements upfront.
The Progressive Reveal: start broad and let each answer shape your next question. Use it for research and brainstorming.
The Adversarial Check: after Claude recommends something, ask: "Now make the strongest case against that recommendation."
Build one more pattern into muscle memory: ask Claude to push back.
"I'm thinking about X. What am I missing?" or "Poke holes in this plan before I commit." Most people paste in an idea and ask Claude to polish it. The more valuable move is asking Claude to challenge it. It surfaces assumptions you made unconsciously and risks you haven't considered. You're not outsourcing your thinking, you're pressure-testing it.
Template: "Here's my current plan: [paste]. What are the weakest assumptions I'm making? What risks am I not accounting for?"
Context turns generic output into useful output. The five elements are who, what, how, why, and what not. Iteration turns good output into great output, using the Layered Brief, Progressive Reveal, and Adversarial Check patterns. Start a new conversation when you switch topics; continue when building on the same project. Next: giving Claude the right context upfront through system prompts.