Smorster

32 Tricks to Level Up Claude Code in 16 Mins

Nate Herk | AI Automation · 16m 16s · Watch on YouTube · 14 sources

Decision Card

Effort: ~20 minutes to watch + cherry-pick: most “hacks” are single slash commands or one-line prompt habits you can adopt in the next session (/init, /context, /compact, plan mode); a few (git worktrees, Context7 MCP install, VPS hosting) are 30–60 minute setups.

Honest take: Solid, accurate beginner-to-intermediate roundup — the slash commands and feature names check out against current Claude Code docs. But it’s a fast anthology with zero depth per item, the host’s “32” list is padded (two tricks are numbered #28, and “treat it like a junior dev,” “make it ask questions,” and “plan mode” are the same idea three times), and several “power user” items (agent teams, the /voice command) are experimental or still rolling out — not the dependable workflows the framing implies.

Concrete next steps (per trick — adopt / try / skip):

  • /context to find token bloat — adopt. Real, ~2 sec, shows exactly what eats your window. Docs [01:49]
  • /init on every project — adopt. Generates a CLAUDE.md project map; standard practice. [00:14]
  • Plan mode before executing — adopt. Highest-leverage habit here; shift+tab to cycle modes. [02:48]
  • Permissions allow/deny instead of --dangerously-skip-permissions — adopt. Deny genuinely overrides allow (docs); safer same-speed autonomy. [14:00]
  • Git worktrees for parallel sessions — try. Real feature (claude --worktree <name>), but only worth it if you actually run 3+ concurrent sessions. Docs [10:35]
  • Context7 MCP for current docs — try. Legit, MIT-licensed, free tier (repo). [15:06]
  • Haiku for sub-agents — try. Sound cost play when sub-agents read large volumes and return summaries. [05:55]
  • /loop for recurring checks — try. Real but expires in ~3 days and only fires while the session is open/idle. Docs [11:57]
  • Agent teams — try (cautiously). Powerful but experimental/disabled by default and more expensive. Docs [14:35]
  • /voice native command — skip for now. Host admits it’s still rolling out; use a separate dictation app meanwhile. [01:16]
  • “95% confidence” prompt phrasing — skip. The percentage is theater; the model can’t measure its own confidence. The underlying habit (make it ask questions) is fine. [03:37]
  • The duplicate “junior dev / ask questions / plan mode” trio — skip the redundancy. Adopt plan mode once and move on. [03:13]

TL;DR

A rapid-fire anthology of ~32 Claude Code tips spanning beginner basics (/init, /context, /compact, plan mode) through power-user features (git worktrees, agent teams, Context7 MCP, VPS hosting). The individual tips are largely accurate against current docs, but the list is padded and shallow — treat it as a checklist to cherry-pick from, not a tutorial.

Key Points

  • Run /init on every project so Claude Code generates a CLAUDE.md “cheat sheet” mapping architecture and conventions 00:14
  • Keep context small and use /context to see exactly what’s consuming tokens — system prompts, files, MCP servers — as percentages 01:49
  • Compact around 60% context (you can tell /compact what to preserve) and /clear between unrelated tasks 02:08
  • Always start in plan mode so Claude researches and outlines an approach before writing code, reducing correction rounds 02:48
  • Deploy sub-agents for parallel work with isolated context windows, and run cheaper models like Haiku for simple/high-volume sub-agent tasks 04:53
  • Keep CLAUDE.md lean (host targets 150–200 lines) and route to separate files for style guides, business context, and reference docs 07:13
  • Use git worktrees (claude --worktree <feature>) to run multiple sessions on isolated branches without overwriting each other 10:35
  • Sometimes hardcode a direct API endpoint instead of an MCP server, since MCP loads all tool definitions into context 11:24
  • “ultrathink” allocates the maximum thinking budget (~32,000 tokens) for hard architecture/debugging decisions 13:34
  • Configure permission allow/deny lists for safe autonomy — deny entries take priority over allow 14:00
  • Install Context7 MCP to inject current, version-specific library documentation before Claude writes code 15:06

Notable Quotes

“The less noise in the context window, the better Cloud performs. It’s simple, but a lot of people ignore this.” 01:46

“Every token that it spends going the wrong direction is just wasted context. So steer tight and steer early.” 07:52

“It is called dangerously skip permissions for a reason. So, the smarter way to go about it is to go into your permissions and explicitly allow the commands that you know are safe.” 14:12

Verified Claims

/context shows a percentage breakdown of what’s consuming your token budget (system prompts, files, MCP servers). 01:49

“ultrathink” allocates roughly the maximum 32,000-token thinking budget. 13:34

Git worktrees let you run parallel sessions via claude --worktree <feature-name> on isolated branches. 10:35

Permission deny rules take priority over allow rules. 14:00

  • Claude Code Docs — Settings
  • Verdict: Confirmed — evaluation order is deny → ask → allow; a deny in any scope cannot be overridden by an allow.

Context7 MCP injects up-to-date, version-specific documentation for popular libraries. 15:06

/loop reruns a prompt on an interval but only lasts ~3 days. 11:57

Custom skills live as markdown files and can be invoked by slash command or natural language. 05:28

  • Claude Code Docs — Extend Claude with skills
  • Verdict: Confirmed, minor correction — skills now live in .claude/skills/<name>/SKILL.md (the video says .claude/skills with a flat .md, which matches the older .claude/commands/ convention that still works).

Agent teams share a task list and communicate directly, unlike sub-agents. 14:35

Tools, Papers & Standards Mentioned

Follow-up Questions

  1. For an agent that downloads, transcribes, and summarizes YouTube videos (this project’s actual goal), which of these tips meaningfully help — e.g., does the Haiku-for-bulk-reading + Opus-for-summary split reduce cost without hurting summary quality?
  2. How do /loop (3-day, session-bound) and desktop scheduled tasks differ in reliability for a “process my researchLater playlist” trigger, and is either appropriate given this project’s “manual trigger only” constraint?
  3. Is hardcoding direct API endpoints over an MCP server (e.g., the YouTube Data API) actually a net token win once you account for the auth/error-handling code Claude has to carry in context?

Sources