Keyboard shortcuts

Claude Code on Windows has a few things outside the docs. Alt+P triggers Plan Mode—you tell the agent “plan this and go,” and it runs without you. Alt+E changes effort (Extra, Max)—sometimes it helps to dial down when you have too many agents running. Alt+T toggles extended thinking—the agent thinks through the problem internally before it answers. I use all three and they cut down on rewriting prompts.

At the start of a new session, /clear is useful (I wipe old memory so the agent starts fresh). /init generates my first CLAUDE.md and suggests improvements if one already exists.

In the browser (Chrome extension) you have Ctrl+K (Mac: Cmd+K) to open commands. There’s “Add to Claude”—you send a photo without typing and the agent sees it right away.

You don’t have to remember all the shortcuts, but Alt+P (plan) and Alt+T (thinking) are worth it.

Lesser-known features

Memory Tool is how the agent remembers things between sessions. It’s not CLAUDE.md (that’s the instructions you give). Memory is a separate system that the agent writes to itself based on your feedback. I use it for things like “George prefers short sentences without commas” or “Euphronic runs on Vercel.” The agent then sees it at the start of a new chat without having to ask you.

Scheduled Tasks: you tell Claude “send me a list of emails and meetings at 8 AM” and it programs that itself. It runs without your supervision. It works up to an hour ahead, but for routines it’s worth it.

Memory Tool and Scheduled Tasks are worth experimenting with.

Prompt tricks

Context first, question second. When you give it everything (inspiration, brand guide, list of sections), the agent picks better. Don’t say “can you build me a website?” Say “here are 5 examples, here’s the brand voice (5 pages), here’s the section list… can you build me a homepage?” The difference is enormous.

Reverse prompting—let the agent ask you questions. Tell it: “ask me about the goal, role, blockers, then generate your own prompt.” The result is more precise than what you’d figure out alone.

Text dividers. Instead of the prompt dumping as one wall, separate sections with two hashes. “## Brand voice,” “## Constraints,” “## Outputs.” The model chews through it better.

Tell it what you don’t know. “I’m not sure this is the best path, but here’s the original file and the history” is 100 times better than pretending you’re certain. The agent will verify it itself then.

Context first, reverse prompting when you’re unsure, dividers when you have multiple parts.

Time savings

Recording instead of typing. When you have a complex task, record it. The Recording feature in Claude Code captures you for 5–7 minutes, the agent transcribes it and returns clean text. Input sounds more natural and the agent picks up your tone.

New chat for a new topic. When you switch gears (design → code bugs), start a new chat. Without it, contexts bleed together and the agent pulls the wrong piece from historical memory. One “Create new chat” takes 3 seconds; a tangled conversation eats hours afterward.

Plan at the start. Tell the agent “plan it, then go” instead of “code it right now.” The plan takes 20 minutes, the rest runs without you. Without a plan the agent cycles and then you have to change things.

Photo instead of text. Instead of writing “change the background color to blue and add a border,” take a screenshot, sketch what you want in the image, and send it. The agent picks out just that piece and changes it.

Recording, new chat, plan, and photo instead of text—these really do save time.

Cost savings

Claude Code costs $100 per month for full Max. Cheaper is $20 for Sonnet alone (slower). I use Max because the agent cycles less. That’s an $80 monthly difference, but when the agent cycles and you spend 10 hours debugging, the economics flip. Pick the price based on what you need today, not what’s cheapest.

Sonnet handles 90 percent. I tried Opus (the pricier model) and learned: when you have time, Sonnet. When you have an hour deadline, Opus. Haiku for things that are straightforward.

The Gemini API once hit me with a thousand-dollar overage. The Free Trial credit I had didn’t cover that API, so it expired useless. Lesson: without a hard cap, autonomous agents will run away. I set up a budget alert in Google Cloud. It takes 5 minutes and saves you surprises.

Vercel is $40 per month for all websites combined. Supabase is $25 for the database. GitHub is free. Cloudflare is free. That’s $65 total. Spread across projects (web, forum, news), it’s nothing each. And because it’s cheap, I don’t spin up unnecessary things.

Sonnet handles 90 percent, hard cap on autonomous agents, simple stacks are cheaper. Verify what yours costs.

From the community

Context folder. Everyone says “context is king,” few do it. In the community I see people create a folder with a brand voice guide and ideal customer, then copy it into every new project. It’s 50 lines of markdown. I do it too. The agent then knows who you are, and the output is completely different.

Reverse prompting: when you’re unsure what you want, let the agent ask you. It takes 5 minutes and the result is more precise.

Mining—things the agent generated that are worth keeping. Before you delete a conversation, download the valuable pieces. Git or Notion. Otherwise it gets lost and later you think “that was great, where is it?” It takes 2 minutes and saves you days.

Validate your brief. When you give the agent a task, it should see the answer already in the brief itself. Poor input equals poor output, regardless of the model.

Context folder, reverse prompting, mining, and clear briefs—these are what sharper people in the community do.