27 Hidden Claude Code Features and Shortcuts Most Users Miss

Adopt a Proper Mental Model

Treat Claude Code as a junior developer that needs clear onboarding instructions, explicit constraints, and periodic calibration. Providing structure lets Claude stay on track and avoid drifting.

Introductory Commands (avoid most ineffective sessions)

Run /init on every new project Creates a CLAUDE.md file that records architecture decisions, conventions, and naming rules. Subsequent sessions automatically load this context. /init Enable /statusline Turns the terminal into a live dashboard showing token usage, session statistics, and model information. /statusline Enable voice input Activate once in settings; then dictate commands without breaking the problem‑solving flow.

Run /context to identify token consumers Shows which files, snippets, or stale commands are using tokens so they can be removed. /context Compress or clear at 60% utilization When the context reaches about 60% of its limit, run /compact to compress the conversation while preserving key information. Use /clear between unrelated tasks.

/compact   # use for long‑running tasks
/clear      # use when switching to a new unrelated task

Intermediate Practices (after mastering basics)

Start in planning mode Ask Claude to outline a plan, list steps, and expose assumptions before writing any code. This adds a calibration point and prevents later rework.

"Before doing anything, list the approach.
Write out steps and note all assumptions.
If anything is unclear, ask me first."

Planning adds ~3 minutes up front but can prevent 3‑hour fixes later.

Describe the problem, not just a command Provide context, constraints, and relevant details (e.g., error description, UI behavior) instead of a terse request like “create a login page.”

Let Claude ask clarification questions first Append a pre‑prompt that asks Claude to list questions, answer them once, then let Claude proceed.

Embed self‑check steps in to‑do lists Include verification steps (e.g., test with empty input, special characters, large amounts) so Claude runs these checks before declaring completion.

Todo:
1. Build payment form
2. Add input validation
3. ✓ Verify with empty input, special chars, $10,000 amount
4. Add error messages
5. ✓ Verify all error states are visible

Save reusable prompts as custom skills Store instruction sets in files such as skills/code-review.md and load them to enforce consistent standards.

# skills/code-review.md
1. Check security vulnerabilities
2. Flag N+1 queries
3. Verify error handling for external calls
4. Ensure camelCase naming
5. Output: list critical issues first, then suggestions

Match model to task complexity Use lightweight models (Haiku) for simple renaming, refactoring, or configuration changes. Use larger models (Sonnet/Opus) for architecture decisions, complex debugging, or multi‑file changes. Overusing Opus is compared to “taking a Ferrari to buy groceries.”

Advanced Practices (turn Claude into a team)

Deploy sub‑agents for parallel work Launch isolated sub‑agents (e.g., one for database schema, another for API endpoints) and coordinate them yourself. This yields multiplicative productivity gains.

Save frequently used prompts as skills Load a skill file so Claude consistently applies the defined standards without rewriting instructions.

Continuously update CLAUDE.md Edit the file as the project evolves to reflect the current tech stack, strict rules, and known issues. An outdated file leads Claude to make confident but wrong decisions.

# CLAUDE.md
## Current Tech Stack
- Next.js 14 (App Router)
- Supabase for auth & DB
- Tailwind + shadcn/ui for components
## Strict Rules
- Never commit or push directly; user handles git
- Always use TypeScript
- All API routes live under /app/api/
## Known Issues
- Mobile auth sessions may expire; always check

Link external files instead of embedding everything Reference external docs in CLAUDE.md (e.g., /docs/architecture.md ) so Claude loads only needed parts, keeping the token budget healthy.

Terminate a diverging session immediately If Claude veers off path, stop the session, rewrite the prompt with a clear restatement of requirements, constraints, and focus, then restart.

Challenge the first output Never accept the initial answer. Ask “What is the weakest part? How would you improve it with more time? Is there a simpler solution?” to drive better results.

Set notification hooks Configure a hook so Claude notifies you when a long task finishes, allowing you to attend to other work.

# Example notification
claude run "build the feature" && \
osascript -e 'display notification "Claude done" with title "Claude Code"'

Use screenshots for visual debugging Paste a screenshot of the UI into the session; Claude can analyze layout issues and suggest CSS fixes.

Integrate Chrome DevTools Ask Claude to open DevTools on a page, read console errors, network requests, and DOM state, turning debugging into a visual, interactive process.

Resume a previous session Use claude --resume <session-id> to reopen a session with full history and context. claude --resume 9901b366-c4d78-89ad-81e964e373e Run parallel sessions with Git worktrees Create separate worktrees for feature branches and run independent Claude sessions in each, delivering a week’s work in a single day.

# Create worktrees
git worktree add ../project-feature-auth feature/auth
git worktree add ../project-feature-payments feature/payments
git worktree add ../project-feature-dashboard feature/dashboard
# Run Claude in each directory simultaneously

Replace MCP server with direct API calls Calling APIs directly reduces overhead, token consumption, and latency compared to routing through an MCP server.

# Bad: via MCP → Notion
# Good: direct call
"Call Notion API: api.notion.com/v1/pages"
"Here is my API key format and endpoint structure."

Automate repetitive tasks with /loop /loop lets Claude repeat a task unattended for up to three days, continuously monitoring, logging, and categorizing results.

Remote control via phone Expose a Claude session with ngrok http 3000 and control it from a mobile browser, allowing you to start builds while away from the desk.

# Expose session
ngrok http 3000
# Access from phone, start build, and let it finish

Query databases with natural language Ask Claude a question; it translates it to SQL, runs the query, and returns the answer—no need to write SQL or remember the schema.

Use Ultrathink for truly hard problems Allocate the maximum token budget and explicitly tell Claude to use Ultrathink for complex architecture decisions, security‑critical implementations, or performance optimizations.

Activate Ultracode mode in Opus 4.8 Run /effort high to enable the highest‑level coding mode, which consumes more tokens but yields superior code for difficult tasks.

/effort high

Progression Plan

Week 1 : Run /init, use /compact at 60%, adopt planning mode, and embed self‑check steps in to‑do lists.

Week 2 : Add custom skills, switch to lightweight models for simple tasks, keep CLAUDE.md up to date, and challenge each output.

Week 3 : Deploy sub‑agents, use Git worktrees for parallel sessions, and run multiple feature branches simultaneously.

Week 4 and beyond : Use /loop for long‑running automation, enable Ultrathink for difficult problems, integrate Chrome DevTools, and control sessions remotely via ngrok.