Smorster

32 Tricks to Level Up Claude Code in 16 Mins

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

Decision Card

Effort: Skim-and-cherry-pick anthology — most tricks are a single slash command, prompt habit, or settings tweak (2–10 minutes each); the heaviest items (git worktrees, agent teams, permission allowlists) are a half-day of setup and practice combined.

Honest take: The video treats fast-moving Claude Code features as stable facts and is already partly stale: “ultrathink” was deprecated in January 2026 (extended thinking is now on by default at the same 31,999-token budget), /voice was only at ~5% rollout when covered, remote control requires a Max-plan research preview, and agent teams sit behind an experimental flag — none of which the video flags. It also numbers two different tricks as “28” (so the count only works in the title) and casually floats having the browser agent solve CAPTCHAs, which is exactly the kind of ToS-violating automation the tooling vendors prohibit.

Concrete next steps:

  • adopt — /init plus a lean CLAUDE.md (150–200 lines) that routes to detail files (tricks 1, 14, 15); ~15 min, guidance at claude.com/blog/using-claude-md-files
  • adopt — context hygiene: /context to diagnose bloat, /compact with keep-instructions around 60%, /clear between unrelated tasks (tricks 4–6); zero setup, pure habit
  • adopt — plan mode first + “keep asking questions until 95% confident” prompting (tricks 7–9); habit only
  • adopt — replace --dangerously-skip-permissions with explicit allow/deny rules (trick 30); ~30 min against code.claude.com/docs/en/permissions
  • try — self-verifying to-do lists plus screenshot / Chrome DevTools MCP check loops for front-end work (tricks 10, 20, 21); ~1 hr with github.com/ChromeDevTools/chrome-devtools-mcp
  • try — claude --worktree parallel sessions (trick 23); ~30 min, docs at code.claude.com/docs/en/worktrees
  • try — Haiku-powered subagents for bulk reading/scraping while the main thread stays on a stronger model (tricks 11, 13); ~20 min
  • try — custom skills in .claude/skills/ for repeated SOPs (trick 12); ~30 min per skill
  • try — /statusline, /rewind, and notification hooks (tricks 2, 18, 19); ~10 min each
  • try — Context7 MCP for current library docs (trick 32); ~10 min via github.com/upstash/context7
  • try — /loop for in-session recurring checks and CLI data analytics against tools like bq (tricks 25, 28a); minutes, but note the 3-day loop expiry
  • skip — typing “ultrathink” (trick 29): deprecated since January 2026; extended thinking is default now
  • skip — the CAPTCHA/form-automation aside in trick 21: ToS and legal exposure, and unreliable anyway
  • skip — VPS hosting and Telegram control (trick 26) unless you already have genuinely long-running unattended jobs; remote control (trick 27) covers the “walk away from your desk” case with less setup, but only on Max plan today
  • skip — wholesale site cloning (trick 22); use screenshots as style inspiration only
  • skip — /voice (trick 3) until the rollout reaches your account; check the welcome screen banner

TL;DR

A 32-item anthology of Claude Code tips spanning beginner habits (CLAUDE.md via /init, context hygiene with /context//compact, plan mode), intermediate workflow (subagents on cheap models, custom skills, screenshot self-check loops, /rewind, hooks) and power-user features (git worktrees, /loop, remote control, agent teams, Context7 MCP, permission allowlists instead of --dangerously-skip-permissions). The individual tips are largely accurate and verifiable against official docs, but several depend on rollout status, plan tier, or experimental flags the video doesn’t mention, and at least one (“ultrathink”) is already obsolete.

Key Points

  • Run /init on every project so Claude Code scans the codebase and generates a CLAUDE.md “cheat sheet” of architecture, conventions, and key files 00:14
  • Manage context deliberately: /context shows what’s eating tokens, /compact (optionally with keep-instructions) compresses history around 60% usage, /clear wipes between unrelated tasks 01:51
  • Start work in plan mode (shift+tab) so Claude researches and outlines before touching code, and prompt it to keep asking questions until 95% confident 02:48
  • Bake verification into the to-do list itself — e.g. “build website” followed by “screenshot it and check the layout” and “open Chrome DevTools and check functionality” 03:58
  • Use subagents for parallel work, running them on Haiku for cheap bulk reading while the main thread stays on Opus 04:54
  • Keep CLAUDE.md between 150–200 lines and have it route to separate files (style guides, business context) rather than inlining everything, because it loads into every conversation 07:06
  • A screenshot self-check loop — design, screenshot, revise, repeat three times before showing V1 — markedly improves front-end first drafts 09:19
  • claude --worktree <feature> gives each session an isolated branch/workspace so multiple sessions on one repo don’t overwrite each other 10:35
  • When token-constrained, hardcode the one or two direct API endpoints you need instead of loading a full MCP server’s tool definitions into context 11:24
  • Prefer explicit allow/deny permission rules over --dangerously-skip-permissions; deny entries override allow entries 14:02
  • Agent teams go beyond subagents: agents share a task list, message each other, assign each other work, and you can address individual teammates directly 14:35
  • Context7 MCP injects current version-specific library documentation before Claude writes code, countering training-data cutoff hallucinations 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.” 07:53

“And anything in the deny list is going to take priority over anything in the allow list.” 14:32

Verified Claims

  • Claim: /init scans the codebase and generates a CLAUDE.md mapping architecture, conventions, and key files. 00:21 Sources: Claude blog — Using CLAUDE.md files, Claude Code best practices Verdict: Confirmed.

  • Claim: Typing “ultrathink” allocates the maximum thinking budget of ~32,000 tokens. 13:44 Sources: ClaudeLog — What is Ultrathink, Decode Claude — UltraThink is Dead Verdict: Disputed — the 31,999-token budget figure is right, but the keyword was deprecated in January 2026; extended thinking at that budget is now the default, so typing the word no longer does anything special.

  • Claim: claude --worktree <name> creates an isolated workspace on its own branch so parallel sessions don’t collide. 10:55 Sources: Claude Code docs — Worktrees Verdict: Confirmed (native --worktree/-w flag; worktrees live under .claude/worktrees/).

  • Claim: Deny-list entries take priority over allow-list entries in permissions. 14:32 Sources: Claude Code docs — Configure permissions Verdict: Confirmed — rules evaluate deny → ask → allow, and a deny at any settings level beats an allow at any other.

  • Claim: /loop recurring tasks only last 3 days. 12:24 Sources: Better Stack — The /loop skill and Desktop scheduler, claudefa.st — Scheduled tasks guide Verdict: Confirmed — loops expire 3 days after creation and don’t persist across session restarts.

  • Claim: Remote control lets you steer a local session from your phone or browser while code never leaves your machine. 12:56 Sources: Claude Code docs — Remote Control Verdict: Confirmed, with an omission — it’s a research preview for Max-plan subscribers, the terminal must stay open and online, and only chat messages and tool results cross the bridge.

  • Claim: A native /voice command exists and is “still rolling out.” 01:18 Sources: Claude Code docs — Voice dictation, TechCrunch — Claude Code rolls out voice mode Verdict: Confirmed — began rolling out March 2026 to ~5% of paid users initially; note audio is transcribed on Anthropic’s servers, not locally.

  • Claim: Agent teams let parallel agents share a task list, message each other, and assign each other work, unlike isolated subagents. 14:38 Sources: Claude Code docs — Agent teams Verdict: Confirmed, with an omission — the feature is experimental and off by default (CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1).

  • Claim: Context7 MCP injects up-to-date, version-specific library documentation into context before Claude writes code. 15:17 Sources: GitHub — upstash/context7, Upstash blog — Context7 MCP Verdict: Confirmed (exposes resolve-library-id and get-library-docs tools; API key optional but recommended).

  • Claim: /rewind rolls back to a previous point in the conversation without starting over. 08:23 Sources: Claude Code docs — Checkpointing Verdict: Confirmed, with a caveat — it can restore code, conversation, or both, but only tracks Claude’s own file-edit tools; changes made via bash commands (rm, mv, cp) are not recoverable.

Tools, Papers & Standards Mentioned

Follow-up Questions

  1. At what point does the token overhead of a loaded MCP server (tool definitions in context) actually exceed the cost of hardcoded API endpoints — can this break-even be measured for a real server like Notion’s?
  2. How do agent teams compare with plain worktree-parallel sessions on cost and output cohesion for a mid-sized feature, given teams are “more expensive and run longer”?
  3. Does the “screenshot three passes before V1” self-check loop measurably beat a single-pass build plus human feedback on front-end quality per token spent?

Sources