9 Claude Code Hacks to Use It Better Than 90% of People

Every Claude Code technique worth using has two things in common: it's high leverage, and it's dead simple to set up. Here are nine of my favorites — the tips, hacks, and workflows that put you ahead of 90% of Claude Code users. Most take under a minute to implement. All of them are immediately useful.

Tip #1: Stop Using MCPs, Start Using CLIs

Anywhere you can replace an MCP server with a CLI, do it. CLIs are more efficient, use fewer tokens, and often give you more functionality.

The reason is simple. Command line interfaces live in the terminal. Claude Code lives in the terminal. There's no protocol overhead, no server to run, no translation layer — just a native tool call. MCPs had their moment, but the ecosystem has been shifting toward CLI tools for the last several months because the numbers don't lie.

One thing to pair with this: when you install a CLI for Claude Code to use, always include a matching set of skills. The CLI is the tool; skills are what teach Claude Code how to use it properly. Those two go together.

Tip #2: Use /btw For Free Sidebar Conversations

The /btw (by the way) command lets you hold a sidebar conversation with Claude Code in the same session without adding to the context window. It's essentially free.

Here's why it matters. Normally, when Claude Code is in the middle of something — say, a long web search — you either wait for it to finish or start a new session. /btw gives you a third option: tap it on the shoulder while it's working, ask a quick question, get an answer with full context, and let it continue what it was doing.

Because the /btw exchange doesn't count against the active context window, it's essentially a free conversation. On long-running tasks with big context, this turns into a legit efficiency hack. Use it for quick clarifications, status checks, or sanity questions without burning context.

Tip #3: Play A Sound When Claude Code Finishes

This one is stupid simple but saves hours a week. Use /hook to set up a sound notification when a task completes.

The problem: if you're like me, you've got three or four Claude Code windows open at any given time. You fire off a task, switch contexts, forget about it, and 20 minutes later you remember it's been waiting for you that whole time. Multiply that across a week and it's real time lost.

The fix: /hook and say "play a sound when you complete a task." You can point it at any audio file you want — I use something small and subtle so it doesn't annoy me. Over the course of a week, this saves genuine hours of context-switching penalty.

Tip #4: Run /clear Early And Often

Claude Code's performance degrades measurably as context fills up. Clear the context window well before you hit the ceiling — ideally around 20-25% of your total tokens.

Even with a 1 million token context window, performance doesn't stay flat. Anthropic's own long-context retrieval data tells the story:

• At 256,000 tokens, retrieval efficiency sits around 91-92%.
• At 1 million tokens, it drops to 78.3% on Opus and 65.1%.

That's a massive quality cliff. If you don't need the context you've accumulated — and most of the time you don't — just /clear and start fresh. Context rot is real. The more you let it build up without reason, the worse Claude Code gets at everything it's doing.

Rule of thumb: if you're past 20-25% of your context window, ask yourself why you're still in that session. If there's no good answer, clear.

Tip #5: Set Up A Status Line

The status line is the bottom bar in your Claude Code terminal that shows your current directory, model, and context window usage. It's one prompt to set up, and once it's running you always know where you stand.

You don't have to run /context to check usage. You don't have to guess which model you're on. You just see it. I keep my status line showing directory, model, and context window percentage — which means I always catch myself when I'm creeping past that 20-25% threshold and should clear.

Setup is a single command: /statusline followed by a prompt describing what you want shown. Takes 30 seconds. Never not useful.

Tip #6: The Skill Creator Skill From Anthropic

The Skill Creator skill, built by Anthropic, does more than create skills — it also modifies, improves, and measures skill performance with actual benchmarks.

Anytime I'm touching a skill, I use Skill Creator. The part people miss: it doesn't just rewrite skills, it benchmarks them. You can run it against a skill, make a change, run it again, and get quantifiable before/after numbers. No more guessing whether your "improvement" actually improved anything.

If you're serious about the skills ecosystem — building your own, customizing others, iterating on prompts — this is table stakes.

Tip #7: Enable Agent Teams For Coordinated Subagents

Agent teams are an experimental Claude Code feature where subagents can talk to each other and coordinate, unlike standard subagents which each work in isolation.

Standard subagents are useful, but they're siloed. You can have three running in parallel, doing three different things, but they only talk to the orchestrating Claude Code session — not to each other. That's fine for independent tasks; it's a limitation for anything that needs coordination.

Agent teams change that. Subagents on a team share context, hand off work, and build on each other's outputs. Higher-quality results, at the cost of slightly more tokens. For anything complex enough that siloed subagents aren't cutting it, it's worth the trade.

How To Enable Agent Teams

1. Go to the Agent Teams documentation page.
2. Copy the URL.
3. Paste it into Claude Code and say "enable agent teams."
4. Once enabled, you have to explicitly invoke a team by saying "create an agent team to do X, Y, and Z." It won't auto-trigger.

Tip #8: Ask Open-Ended Questions In Plan Mode

This is the simplest tip on the list and the most underrated. In plan mode, don't just ask Claude Code to do something — ask it what you're not thinking about.

The biggest trap for non-technical users in Claude Code is classic: you don't know what you don't know. Plan mode asks decent surface-level questions by default, but they're safe questions — framework choice, styling preferences, basic scope.

The unlock: explicitly prompt for deeper thinking.

• "What am I not thinking about?"
• "What are the unintended consequences of this approach?"
• "What would an expert in [domain] be considering here?"
• "What are the failure modes you'd expect?"

Do this consistently and you get two things: dramatically better outputs, and you educate yourself on what experts actually worry about. Over time you learn to ask the right questions yourself. Until then, let Claude Code flag what you'd otherwise miss.

Tip #9: Use Claude Code + Obsidian As A Second Brain

Claude Code is excellent as a personal assistant — especially when paired with Obsidian, a free markdown-based knowledge management app that integrates cleanly with the terminal.

The setup is trivial:

1. Download Obsidian.
2. Create a vault (just a folder on your computer designated as the Obsidian vault).
3. Open Claude Code inside that folder.

That's it. Now every markdown file Claude Code creates shows up in Obsidian with proper linking and the graph view. You navigate your notes through Obsidian's UI; Claude Code stays organized because the structure is visible to it. The relationship is symbiotic: Obsidian gives you a UI over Claude Code's output, Claude Code gives Obsidian a writer.

This doesn't replace traditional code projects — it's not meant to. It's for any project where the artifact is a growing pile of markdown: research, notes, playbooks, personal assistant workflows, content planning. Anywhere you'd use a wiki or a second brain.

FAQ

Why should I stop using MCPs?

CLIs outperform MCPs for most use cases because they run directly in the terminal without protocol overhead. They use fewer tokens, tend to be more reliable, and are easier to debug because you can call them yourself. MCPs still make sense for some niche cases but the default preference has shifted to CLIs.

How much does context rot actually hurt performance?

At 1 million tokens of context, Claude Opus retrieval efficiency drops to 78.3%, and non-Opus models drop further to 65.1%. That's compared to roughly 91-92% at 256K tokens. For any task that requires accurate recall of earlier content, staying under 25% of your context window is a concrete quality win.

What's the difference between a regular subagent and an agent team?

Regular subagents operate in isolation — they only communicate with the orchestrating session, not with each other. Agent teams allow subagents to talk to each other, coordinate work, and share context. Agent teams are still experimental and must be explicitly enabled and invoked.

Do I need to be a developer to use Claude Code with Obsidian?

No. That's actually the whole point. Obsidian + Claude Code works best for non-code use cases — personal knowledge management, research, note organization, content workflows. If you're not a developer, this is probably the highest-leverage way to start using Claude Code as a personal assistant.

What should my status line show?

At minimum: current directory, current model, and context window usage percentage. Directory prevents you from accidentally running commands in the wrong folder. Model tells you whether you're on Opus, Sonnet, or Haiku. Context percentage tells you when it's time to /clear.

---

If you want to go deeper on Claude Code workflows, join the free Chase AI community for templates, prompts, and live breakdowns. And if you're serious about building with AI, check out the paid community, Chase AI+, for hands-on guidance on how to make money with AI.