Glimpse.
Nate Herk | AI Automation
16 min video
3 min read
32 Claude Code Hacks: Beginner to Power User
You just saved 13 min.
The big takeaway
Master Claude Code through 32 progressive tricks spanning initialization and context management (beginner), parallel agents and custom skills (intermediate), and advanced features like work trees, remote control, and real-time documentation integration (expert). Each tier unlocks faster workflows, lower token costs, and higher-quality outputs.
Beginner Hacks: Foundation & Context
Run /init on every project
Type /init to scan your entire codebase and auto-generate a cloud.md file that maps architecture, conventions, and key files. This eliminates re-explaining your project each session and lets Claude contextualize everything automatically.
Set up a status line
Use /status line to display a mini dashboard at the bottom of your terminal showing model, context percentage, and cost in real time. This helps you monitor and avoid context rot during long sessions.
Use voice input
Claude Code now has a native /voice command to dictate code directly into the terminal. Voice input is also available via third-party apps for dictation anywhere on your system.
Keep context small
Only feed Claude what it needs for the current task; break big problems into focused steps. Less noise in the context window means better performance and fewer wasted tokens.
Use /context to find token bloat
Run /context to see exactly what is consuming tokens—system prompts, file contents, MCP servers—broken down by percentage. Diagnose and restructure to eliminate waste.
Compact at 60% and clear between tasks
When context hits 60%, type /compact to compress conversation history while preserving critical info (e.g., 'keep all API decisions'). Use /clear to wipe the slate when switching tasks entirely, but retain cloud.md and files.
Always start in plan mode
Hit Shift+Tab to enter plan mode, where Claude reads and researches but does not change anything. It outlines steps, asks clarifying questions, and maps the approach before writing code—proven to improve quality and reduce revisions.
Treat Claude like a junior developer
Instead of direct commands ('Write a function that does X'), pose problems ('How should we handle growth tracking?'). Let Claude reason through decisions, ask it to explain, and you'll get better outputs.
Make Claude ask questions
Tell Claude to invoke its ask-user-question tool and ask continuously until 95% confident it understands your needs. This alignment prevents multiple rounds of revision.
Build self-checking into to-do lists
Embed verification steps directly into execution plans—e.g., after 'build website,' add 'take screenshot and verify layout' and 'open DevTools to check for errors.' This bakes quality checks into the workflow so Claude validates before handing off.
Intermediate Hacks: Speed & Efficiency
Deploy sub-agents for parallel work
Tell the main session to use sub-agents; Claude spins up isolated agents with fresh context windows, each potentially using different models, working in parallel. They research, write tests, and explore approaches independently, then report back to the main agent—like having a team of developers.
Build custom skills
Create reusable prompt files in .cloud/skills/ (e.g., techdebt.md, codereview.md) that define workflows. Invoke them in natural language or via /command to run consistently every time. Commit to GitHub for team-wide reuse.
Use Haiku for sub-agents
Set sub-agents to use Haiku for simple tasks or data processing. Have Haiku scrape articles and summarize for Opus (main agent) instead of wasting expensive tokens on heavy reading—cuts costs without sacrificing quality where it matters.
Haiku (sub-agents, simple tasks)
1 relative cost
Opus (main agent, decisions)
5 relative cost
Cost efficiency by model assignment
Constantly refresh cloud.md
Update cloud.md whenever you discover new patterns, create skills, or find gotchas. This ensures Claude knows your project better next session and prevents repeat mistakes. Keep it lean: 150–200 lines max to avoid context bloat.
Have cloud.md route to other files
Keep cloud.md lean by linking to separate files for style guides, business context, and reference docs. Claude knows where to look without wasting tokens loading info it doesn't always need.
Exit early and re-ask
If Claude goes down the wrong path, hit Escape, correct course, and re-prompt immediately. Every token spent in the wrong direction is wasted context.
Challenge outputs aggressively
Push back on mediocre outputs: 'Scrap that, do a more elegant version' or 'Try a completely different approach.' Claude often delivers dramatically better results on the second try when you set a higher bar. Update skills or cloud.md so it doesn't repeat the mistake.
Use /rewind for quick undos
If you make a wrong turn, type /rewind to roll back to a previous point in the conversation without starting over—fast and clean.
Use hooks for notifications
Type /hooks to set up notification alerts when Claude finishes a session. This lets you run 15+ parallel sessions and get notified when each one needs input, so you can work on other tasks.
Use screenshots for visual feedback
Claude can see and analyze screenshots. Feed it error messages, inspiration sites, or have it screenshot your work and verify layout/functionality. Build a loop: design → screenshot → analyze → implement changes → repeat 3 times before V1.
Use Chrome DevTools for app testing
Claude can open browsers, interact with apps, check functionality, fill forms, and navigate UI. This is huge for front-end work and for automating tasks where no explicit API exists.
Clone inspiration sites
Screenshot sites you like and tell Claude 'Make it look like this.' Claude recreates design patterns without generic AI slop. You can also feed HTML/CSS for styling reference, then add your own touch.
Advanced Hacks: Power User Territory
Run parallel sessions with Git work trees
Use Claude --work-tree [feature-name] to create isolated workspaces on separate branches. Run 3–5 sessions on the same project simultaneously without agents overwriting each other. Merge branches back together when done.
Use API endpoints instead of MCP servers
MCP servers load all tool definitions into context, consuming tokens. For specific tasks, hardcode direct API endpoints instead. Example: if you only need to read one Notion database, skip the full MCP and use just that endpoint.
Use /loop for recurring tasks
Type /loop to rerun a prompt every N minutes (e.g., 'every 5 minutes check deployment'). Claude monitors in the background and interrupts only when attention is needed. Loops last up to 3 days; for longer automation, use desktop scheduled tasks (though each spawn is a fresh session).
Host on a VPS for always-on sessions
Run Claude Code on a remote server so it stays active even when your laptop closes. SSH in anytime to interact, or control via Telegram. Perfect for long-running tasks you don't want to babysit.
Use remote control from your phone
Start a task at your desk, then steer it from your phone or any browser via remote connection. Code stays local; only the connection is remote. Start something heavy, grab coffee, and keep building from your pocket.
Use SQL data analytics without SQL
Connect CLI tools like BigQuery's BQ tool to Claude Code. Ask plain-English questions ('What were our top 10 revenue sources last quarter?') and Claude translates to the right query, runs it, and returns the answer.
Use ultra think for hard problems
Type 'ultra think' to allocate maximum thinking budget (~32,000 tokens) before Claude responds. Use for architecture decisions, complex debugging, big refactors, or when initial attempts aren't working. Don't use for simple fixes.
32,000
tokens allocated for ultra think
Maximum thinking budget per response
Edit permissions for safe autonomy
Instead of using dangerously-skip-permissions, explicitly allow safe commands and deny destructive ones (delete, remove). Deny list takes priority. Get the same speed and autonomy without the danger.
Use agent teams
Unlike sub-agents (isolated, no cross-talk), agent teams share a task list, communicate with each other, and assign work. You can talk to individual agents directly. More expensive and slower, but delivers cohesive output for big projects.
Install Context 7 MCP for live documentation
Install the Context 7 MCP server to pull up-to-date, version-specific technical docs for thousands of libraries (Next.js, React, MongoDB, etc.) before Claude writes code. Fixes the problem of Claude suggesting deprecated or renamed APIs due to training cutoff.
Quick Reference: Hack Progression
32 Hacks Organized by Skill Level
The video progresses through 32 techniques: 10 beginner (init, status line, voice, context management, plan mode, self-checking), 12 intermediate (sub-agents, skills, Haiku, hooks, screenshots, DevTools, inspiration cloning), and 10 advanced (work trees, API endpoints, loops, VPS hosting, remote phone control, SQL analytics, ultra think, permissions, agent teams, Context 7 MCP).
1
Beginner Hacks
10 tricks
2
Intermediate Hacks
12 tricks
3
Advanced Hacks
10 tricks
Distribution of 32 Claude Code techniques
Worth quoting
"The less noise in the context window, the better Cloud performs."
— Nate Herk, at [1:31]
"Every token that it spends going the wrong direction is just wasted context."
— Nate Herk, at [7:33]
"Context 7 fixes that because it has up to date version-specific technical documentation about live code examples from thousands of popular libraries."
— Nate Herk, at [15:06]
Try this
Run /init on your next project to auto-generate cloud.md and establish baseline context.
Set up a /status line to monitor context percentage and token usage in real time.
Try /voice or a third-party voice app to dictate code instead of typing.
Create 2–3 custom skills in .cloud/skills/ for recurring workflows (e.g., code review, tech debt scan).
Use /compact at 60% context and /clear when switching tasks to optimize token usage.
Start your next complex task in plan mode (Shift+Tab) before executing.
Build verification steps into your to-do lists (screenshot, DevTools check, etc.).
Test /loop to monitor a deployment or error log every 5 minutes.
Install Context 7 MCP to access live documentation for your tech stack.
Try ultra think on your next architecture decision or debugging challenge.
Set explicit allow/deny permissions instead of using dangerously-skip-permissions.
Experiment with sub-agents on a parallel task using Haiku for cost savings.
Clone an inspiration website design and feed it to Claude as a reference.
Use Git work trees to run 2–3 parallel Claude Code sessions on the same project.
Made with Glimpse by Wozart
glimpse.wozart.com/v/p48skvv6
Share this infographic
Read this infographic as text

32 Claude Code Hacks: Beginner to Power User

Summary of the video “32 Tricks to Level Up Claude Code in 16 Mins” by Nate Herk | AI Automation.

Master Claude Code through 32 progressive tricks spanning initialization and context management (beginner), parallel agents and custom skills (intermediate), and advanced features like work trees, remote control, and real-time documentation integration (expert). Each tier unlocks faster workflows, lower token costs, and higher-quality outputs.

Beginner Hacks: Foundation & Context

Run /init on every project

Type /init to scan your entire codebase and auto-generate a cloud.md file that maps architecture, conventions, and key files. This eliminates re-explaining your project each session and lets Claude contextualize everything automatically.

Set up a status line

Use /status line to display a mini dashboard at the bottom of your terminal showing model, context percentage, and cost in real time. This helps you monitor and avoid context rot during long sessions.

Use voice input

Claude Code now has a native /voice command to dictate code directly into the terminal. Voice input is also available via third-party apps for dictation anywhere on your system.

Keep context small

Only feed Claude what it needs for the current task; break big problems into focused steps. Less noise in the context window means better performance and fewer wasted tokens.

Use /context to find token bloat

Run /context to see exactly what is consuming tokens—system prompts, file contents, MCP servers—broken down by percentage. Diagnose and restructure to eliminate waste.

Compact at 60% and clear between tasks

When context hits 60%, type /compact to compress conversation history while preserving critical info (e.g., 'keep all API decisions'). Use /clear to wipe the slate when switching tasks entirely, but retain cloud.md and files.

Always start in plan mode

Hit Shift+Tab to enter plan mode, where Claude reads and researches but does not change anything. It outlines steps, asks clarifying questions, and maps the approach before writing code—proven to improve quality and reduce revisions.

Treat Claude like a junior developer

Instead of direct commands ('Write a function that does X'), pose problems ('How should we handle growth tracking?'). Let Claude reason through decisions, ask it to explain, and you'll get better outputs.

Make Claude ask questions

Tell Claude to invoke its ask-user-question tool and ask continuously until 95% confident it understands your needs. This alignment prevents multiple rounds of revision.

Build self-checking into to-do lists

Embed verification steps directly into execution plans—e.g., after 'build website,' add 'take screenshot and verify layout' and 'open DevTools to check for errors.' This bakes quality checks into the workflow so Claude validates before handing off.

Intermediate Hacks: Speed & Efficiency

Deploy sub-agents for parallel work

Tell the main session to use sub-agents; Claude spins up isolated agents with fresh context windows, each potentially using different models, working in parallel. They research, write tests, and explore approaches independently, then report back to the main agent—like having a team of developers.

Build custom skills

Create reusable prompt files in .cloud/skills/ (e.g., techdebt.md, codereview.md) that define workflows. Invoke them in natural language or via /command to run consistently every time. Commit to GitHub for team-wide reuse.

Use Haiku for sub-agents

Set sub-agents to use Haiku for simple tasks or data processing. Have Haiku scrape articles and summarize for Opus (main agent) instead of wasting expensive tokens on heavy reading—cuts costs without sacrificing quality where it matters.

Constantly refresh cloud.md

Update cloud.md whenever you discover new patterns, create skills, or find gotchas. This ensures Claude knows your project better next session and prevents repeat mistakes. Keep it lean: 150–200 lines max to avoid context bloat.

Have cloud.md route to other files

Keep cloud.md lean by linking to separate files for style guides, business context, and reference docs. Claude knows where to look without wasting tokens loading info it doesn't always need.

Exit early and re-ask

If Claude goes down the wrong path, hit Escape, correct course, and re-prompt immediately. Every token spent in the wrong direction is wasted context.

Challenge outputs aggressively

Push back on mediocre outputs: 'Scrap that, do a more elegant version' or 'Try a completely different approach.' Claude often delivers dramatically better results on the second try when you set a higher bar. Update skills or cloud.md so it doesn't repeat the mistake.

Use /rewind for quick undos

If you make a wrong turn, type /rewind to roll back to a previous point in the conversation without starting over—fast and clean.

Use hooks for notifications

Type /hooks to set up notification alerts when Claude finishes a session. This lets you run 15+ parallel sessions and get notified when each one needs input, so you can work on other tasks.

Use screenshots for visual feedback

Claude can see and analyze screenshots. Feed it error messages, inspiration sites, or have it screenshot your work and verify layout/functionality. Build a loop: design → screenshot → analyze → implement changes → repeat 3 times before V1.

Use Chrome DevTools for app testing

Claude can open browsers, interact with apps, check functionality, fill forms, and navigate UI. This is huge for front-end work and for automating tasks where no explicit API exists.

Clone inspiration sites

Screenshot sites you like and tell Claude 'Make it look like this.' Claude recreates design patterns without generic AI slop. You can also feed HTML/CSS for styling reference, then add your own touch.

Advanced Hacks: Power User Territory

Run parallel sessions with Git work trees

Use Claude --work-tree [feature-name] to create isolated workspaces on separate branches. Run 3–5 sessions on the same project simultaneously without agents overwriting each other. Merge branches back together when done.

Use API endpoints instead of MCP servers

MCP servers load all tool definitions into context, consuming tokens. For specific tasks, hardcode direct API endpoints instead. Example: if you only need to read one Notion database, skip the full MCP and use just that endpoint.

Use /loop for recurring tasks

Type /loop to rerun a prompt every N minutes (e.g., 'every 5 minutes check deployment'). Claude monitors in the background and interrupts only when attention is needed. Loops last up to 3 days; for longer automation, use desktop scheduled tasks (though each spawn is a fresh session).

Host on a VPS for always-on sessions

Run Claude Code on a remote server so it stays active even when your laptop closes. SSH in anytime to interact, or control via Telegram. Perfect for long-running tasks you don't want to babysit.

Use remote control from your phone

Start a task at your desk, then steer it from your phone or any browser via remote connection. Code stays local; only the connection is remote. Start something heavy, grab coffee, and keep building from your pocket.

Use SQL data analytics without SQL

Connect CLI tools like BigQuery's BQ tool to Claude Code. Ask plain-English questions ('What were our top 10 revenue sources last quarter?') and Claude translates to the right query, runs it, and returns the answer.

Use ultra think for hard problems

Type 'ultra think' to allocate maximum thinking budget (~32,000 tokens) before Claude responds. Use for architecture decisions, complex debugging, big refactors, or when initial attempts aren't working. Don't use for simple fixes.

Edit permissions for safe autonomy

Instead of using dangerously-skip-permissions, explicitly allow safe commands and deny destructive ones (delete, remove). Deny list takes priority. Get the same speed and autonomy without the danger.

Use agent teams

Unlike sub-agents (isolated, no cross-talk), agent teams share a task list, communicate with each other, and assign work. You can talk to individual agents directly. More expensive and slower, but delivers cohesive output for big projects.

Install Context 7 MCP for live documentation

Install the Context 7 MCP server to pull up-to-date, version-specific technical docs for thousands of libraries (Next.js, React, MongoDB, etc.) before Claude writes code. Fixes the problem of Claude suggesting deprecated or renamed APIs due to training cutoff.

Quick Reference: Hack Progression

32 Hacks Organized by Skill Level

The video progresses through 32 techniques: 10 beginner (init, status line, voice, context management, plan mode, self-checking), 12 intermediate (sub-agents, skills, Haiku, hooks, screenshots, DevTools, inspiration cloning), and 10 advanced (work trees, API endpoints, loops, VPS hosting, remote phone control, SQL analytics, ultra think, permissions, agent teams, Context 7 MCP).

Notable quotes

The less noise in the context window, the better Cloud performs. — Nate Herk
Every token that it spends going the wrong direction is just wasted context. — Nate Herk
Context 7 fixes that because it has up to date version-specific technical documentation about live code examples from thousands of popular libraries. — Nate Herk

Action items

  • Run /init on your next project to auto-generate cloud.md and establish baseline context.
  • Set up a /status line to monitor context percentage and token usage in real time.
  • Try /voice or a third-party voice app to dictate code instead of typing.
  • Create 2–3 custom skills in .cloud/skills/ for recurring workflows (e.g., code review, tech debt scan).
  • Use /compact at 60% context and /clear when switching tasks to optimize token usage.
  • Start your next complex task in plan mode (Shift+Tab) before executing.
  • Build verification steps into your to-do lists (screenshot, DevTools check, etc.).
  • Test /loop to monitor a deployment or error log every 5 minutes.
  • Install Context 7 MCP to access live documentation for your tech stack.
  • Try ultra think on your next architecture decision or debugging challenge.
  • Set explicit allow/deny permissions instead of using dangerously-skip-permissions.
  • Experiment with sub-agents on a parallel task using Haiku for cost savings.
  • Clone an inspiration website design and feed it to Claude as a reference.
  • Use Git work trees to run 2–3 parallel Claude Code sessions on the same project.

More like this