The creator of Claude Code just revealed 7 secrets to using Claude Code (Opus 4.7)
TL;DR
Auto mode is the real unlock for Claude Code — Alex Finn says the new CLI-only auto mode (toggled with Shift+Tab) finally fixes Claude Code’s habit of asking for permission every few seconds, letting it run for hours without constant supervision.
Front-load everything into the first prompt with Opus 4.7 — borrowing from Boris Cherny’s advice, Alex shows that Opus 4.7 is much better at following dense instructions, so instead of drip-feeding steps he dumps the full spec for a Linear-style Next.js app up front.
A single prompt can get you to a surprisingly polished v1 — his one-shot prompt produces a working Linear clone with projects, tasks, statuses, priorities, assignees, and tags, and Alex’s reaction is basically: this looks “pretty much exactly like Linear.”
/recap and notifications turn Claude Code into an async teammate — after long sessions, /recap gives a one-line summary of what the agent did, and notifications can ping you when work is finished so you can stop babysitting the tool.
Use /effort strategically, especially if you’re on the $200/month plan — Alex recommends max effort for the big kickoff prompt, extra high for most follow-up work, and only dropping to high/medium/low for cheap-plan users or tiny tweaks like button colors.
Adaptive thinking has replaced the old extended-thinking toggle — with Opus 4.7, you now control depth through prompting itself: say “think carefully step by step” for a hard feature like collaboration, or explicitly ask it to respond quickly for simple edits.
The Breakdown
Boris Cherny’s 7 tips, but actually applied
Alex opens with a promise not to do the usual YouTube thing of reading Boris Cherny’s X thread and blog post back word-for-word. Instead, he wants to prove each tip live by building an app in Claude Code, framing the whole video as a practical “let’s be about that action” session rather than commentary.
Auto mode fixes Claude Code’s most annoying habit
He switches from Claude Desktop to the Claude Code CLI because some of the newest features only exist there right now. The first move is enabling auto mode with Shift+Tab, which he calls a massive upgrade because Claude stops asking for permission every five seconds and can make its own judgment calls about what actions are safe.
The big prompt strategy: build a Linear clone in one shot
The second tip is front-loading as much information as possible before the session starts. Instead of prompting step by step, Alex gives Claude one dense brief: build a Linear-style project management tool in Next.js with projects, tasks, issues, statuses, priorities, assignees, and any extra useful features it thinks belong.
The app ships fast, and it looks legit
Once the run finishes, he pulls up the result and is clearly impressed: a clean interface with project-linked tasks, task creation, statuses, assignees, and tags. His demo task is “finish this YouTube video,” and the main takeaway is that a beginner could walk away with a usable project management app from a single well-structured prompt.
When to use Claude Code vs OpenClaw
In the middle of the build, Alex answers a question he gets constantly: when should you use Claude Code instead of OpenClaw? His split is simple — OpenClaw is for quick prototypes and on-the-go building in Telegram, while Claude Code is for deeper, more complex coding sessions and consumer-facing apps.
/recap and /effort make long sessions manageable
The third tip is /recap, which generates a one-line summary of everything the agent has done so far — especially useful if you’ve let Claude run overnight and need to catch up fast. Then he shows /effort, arguing that most people on Anthropic’s $200/month plan should use max for the initial giant prompt and extra high for almost everything after that.
Notifications turn Claude into an employee you can trust
From Boris’s blog post, Alex highlights notifications as another trust-building feature: you can literally tell Claude Code to set up notifications after each piece of work. He loves this because Claude loads the needed skill and hook itself, then pings you when a long task is done so you can go do something else and come back only when it matters.
Adaptive thinking is now a prompt, not a toggle
The last tip is about the replacement for extended thinking in Opus 4.7. Instead of flipping a setting, you now steer thinking depth through language — “think carefully step by step” for a tricky feature like collaborative mode, or “prioritize responding quickly” for something trivial like changing a button color — and Alex explains this as controlling the thinking budget per task rather than overall effort.