Essential Claude Code Tips: Target Your Biggest Pain Point and Try One Shortcut Tomorrow

1. Basics

Alias : alias cc='claude --dangerously-skip-permissions' added to ~/.zshrc to bypass permission prompts.

! Prefix : Use !git status to run a bash command directly and have the result automatically enter the context.

Esc Stop / Esc×2 Backtrack : Press Esc once to stop the current action; press twice to open a checkpoint menu for restoring code, dialogue, or both. Use --continue to resume the previous session or --resume to select a historic session.

Ctrl+S Stash Prompt : Temporarily save an unfinished prompt, submit a quick question, and later automatically restore the draft.

Ctrl+B Background Run : Send long commands (e.g., tests or builds) to the background with Ctrl+B while Claude continues working.

/clear : Clear unrelated tasks before switching context; accumulated context can degrade output quality.

/btw : Pop up a floating layer for quick questions without polluting the main conversation.

/insight : Generate an HTML report analyzing the past month’s usage patterns, suggesting custom commands and skills.

2. Prompt Techniques

Self‑check Prompt : Include test commands or lint checks in the prompt so Claude runs and fixes them, improving quality 2‑3×.

Paste Raw Data, No Bug Explanation : Directly paste error logs or CI output with "fix"; also pipe input, e.g., cat error.log | claude "fix this".

@ File Specification : Use @src/auth/middleware.ts to locate a file instantly, saving search effort.

Fuzzy Prompt for Code Exploration : Ask "What would you improve in this file?" for unfamiliar repositories.

Let Claude Interview You : Say "Interview me about [feature]"; Claude asks details, boundaries, trade‑offs, then produces a complete spec.

Two Failed Corrections → /clear : If Claude fails to fix after two attempts, clear the context and rewrite the prompt to avoid poisoned context.

3. Thinking & Planning

"ultrathink" Keyword : Triggers Opus deep reasoning for architecture decisions and complex debugging; set persistent thinking level with /effort.

Plan Mode (Shift+Tab Switch) : When facing multi‑file changes or uncertain solutions, plan first; skip small tasks.

/model opusplan : Switch to the Opus plan model.

Ctrl+G Edit Plan : Directly modify Claude’s proposed plan steps in the editor.

4. Context Management

1M Token Context : Switch to a 1‑million‑token model with /model opus[1m] or sonnet[1m]. Control compression trigger timing via CLAUDE_CODE_AUTO_COMPACT_WINDOW.

Guided Compression : Use /compact focus on API changes and modified files or embed persistent compression commands in CLAUDE.md.

Subagent : Command "use subagent to investigate payment flow" to create an isolated context instance, investigate, and return a summary while keeping the main session clean.

/branch (or /fork) : Duplicate the current conversation to explore an alternative path without affecting the original.

5. Parallelism & Collaboration

--worktree feature-auth : Create an isolated worktree and new branch, allowing 3‑5 Claude sessions to run concurrently.

Agent Teams (Experimental) : Enable CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS so Claude distributes tasks among multiple teammates for parallel execution, avoiding simultaneous edits on the same file.

claude -p Batch Fan‑out : Loop with & to process a file list in parallel, limiting tools via --allowedTools.

One Write, One Review : Session A implements a feature, Session B performs a fresh‑context code review; alternatively, A writes tests while B writes the implementation.

/rename + /color : Name and color‑code terminals to differentiate multiple sessions.

6. CLAUDE.md and Rules

/init Trim Half : Auto‑generated content tends to be verbose; keep only essential parts.

Per‑Line Validation : Ask "Will Claude fail without this line?"; if not, delete it. Budget ~150‑200 lines; system prompts already occupy ~50 lines.

Post‑Error Rule Update : After an error, instruct Claude to "update CLAUDE.md to avoid recurrence," letting Claude write its own evolving rules.

.claude/rules/ Directory : Split rules by theme; add paths front‑matter to load conditionally based on file type.

@imports : Reference external docs, e.g., @docs/git-instructions.md, without inflating the main file.

Skills (.claude/skills/) : Load domain‑specific knowledge on demand, more precise than CLAUDE.md.

7. Hooks (Deterministic Execution)

CLAUDE.md provides suggestions (~80% adherence); Hooks are enforced (100% adherence).

PostToolUse Auto‑format : After each file edit, Claude automatically runs Prettier/ESLint.

PreToolUse Dangerous Command Intercept : Blocks patterns like rm -rf, drop table, truncate directly.

Notification Hook Preserve Compressed Context : After compaction, automatically reinject current task, modified file list, and hard constraints.

Stop Hook Play Sound : Claude plays a system sound upon completion, aiding screen‑switching work.

8. Tools & Plugins

LSP Plugin (Highest Priority) : Install with /plugin install typescript-lsp@claude-plugins-official to auto‑diagnose type errors and unused imports.

gh CLI : Uses less context than MCP server for handling PRs, Issues, CI.

Teach Claude Any CLI : Provide a command like sentry-cli --help for Claude to learn usage and then query recent production errors.

/permissions : Whitelist trusted commands to skip repeated confirmations.

/sandbox : OS‑level isolation; writes limited to project directory, network limited to whitelisted domains; use Docker for long‑running unattended tasks.

9. Others

Remote Control : Start with claude remote-control; control via phone or browser at claude.ai/code.

Voice Input : Enable with /voice; hold space to dictate, mixing voice and typing for richer context.

/loop 5m Periodic Check : Poll deployment status or CI pipelines in the background; expires after 3 days.

/statusline : Generate a script showing branch, directory, and context usage rate in the status bar.

/config Output Style : Choose Explanatory, Concise, or Technical, or create custom styles under ~/.claude/output-styles/.

Custom Spinner Text : Replace spinner verbs with fun phrases (e.g., Harry Potter spells) for entertainment.

Conversational PR Review : Ask "What is the most dangerous change in this PR? Will it crash under concurrency?" for deeper insight than a single static review.

10. Supplement

Auth / Payment / Data Changes Must Be Human‑Reviewed : Regardless of automation, these decisions should never be fully delegated to AI.

No Need to Use Everything : Pick the most painful point and try a single tip; that’s sufficient.