Claude Code 15 个实用技巧：从够用到真正好用

I’ve been using Claude Code for over half a year. Hit plenty of walls, and gradually worked out what actually makes it good.

Here are 15 tips. All from real use. No “practice more and you’ll get better” filler.

Context Management

1. CLAUDE.md Is Infrastructure, Not a Notebook

A lot of people treat CLAUDE.md like a notepad, throwing everything into it. The file gets longer and longer, which actually dilutes the important information.

The correct use: CLAUDE.md only holds constraints and pointers, not content.

# Project Rules
- Must run npm run build before deploying
- Never modify published article slugs

# Where to Find Things
- Architecture decisions → memory/knowledge/
- Last session's progress → memory/logs/latest-date_daily.md
- Deployment procedure → memory/skills/deploy.md

Keep it under 50 lines. Details live in the memory layers. CLAUDE.md just points to them.

2. Use /clear Instead of Opening a New Window

When the context window fills up, most people close the terminal and start over. The right move is /clear — it wipes the conversation history while keeping the context loaded from CLAUDE.md.

Opening a new window re-reads CLAUDE.md and achieves the same thing, but /clear is faster.

3. Use # to Reference Files, Not “Go Look at That File”

❌ "Go look at src/components/SearchDialog.tsx"
✅ "Fix the search results layout in #src/components/SearchDialog.tsx"

The # prefix makes Claude read the file immediately, instead of waiting for it to decide whether or not to read it. When context is limited, that difference matters.

4. Write a Plan Before Starting Long Tasks

For tasks that take more than 3 steps, have Claude write an execution plan first and confirm it before starting:

Give me an implementation plan first. Don't start writing code yet.
I'll say "go ahead" when the plan looks right.

This catches misunderstandings before Claude generates a wall of code in the wrong direction.

Workflow Acceleration

5. /loop Automates Repetitive Tasks

Claude Code supports looped execution:

/loop 30m "Check tasks.yaml backlog, handle one in_progress item"

Good for: scheduled sync, heartbeat checks, batch content processing. Not good for: complex tasks that need real-time judgment.

6. Use Task Files Instead of Long Conversations

After 10+ rounds, a conversation’s information density drops and Claude starts “forgetting” things. A better approach:

Write the task into a markdown file and read it each session:

Read tasks/current-sprint.md, handle the first incomplete item by priority.
After completing it, mark the status in the file, then report back.

Task state persists in the file, not in conversation history.

7. Use --dangerously-skip-permissions to Unlock Speed (With Caution)

In well-trusted projects, having to confirm every tool call significantly slows things down.

In controlled environments (projects with Hooks configured), you can skip permission prompts:

claude --dangerously-skip-permissions

Prerequisite: You must have Hooks configured as a safety net first. Otherwise you’re flying without instruments.

8. Use Git Worktree for Experimental Changes

When you’re not sure Claude’s changes are correct, isolate with git worktree:

git worktree add ../experiment-branch -b experiment
cd ../experiment-branch
# Let Claude make changes here
# Merge if it works out, delete if it doesn't

Claude Code handles worktrees well — it correctly identifies the current working directory.

Prompt Techniques

9. Give Claude a Role, Not Just a Task

❌ "Refactor this function"
✅ "You're a Python performance engineer. Refactor this function with a focus on reducing unnecessary memory allocations"

A role gives Claude a judgment framework. Output quality noticeably improves.

10. Constrain with “Don’t Do” Rather Than “Only Do”

Claude tends to “help with more” — but sometimes you just want the minimum change:

Only modify the error handling logic in the getUser function.
Don't refactor other functions, don't add type annotations, don't change the function signature.

Explicitly saying what not to do is more effective than saying “just do this one thing.”

11. Ask Claude to Explain Before Modifying

For important files, have Claude explain the existing code logic before proposing changes:

Explain how the verify_token function in auth.py works,
especially the token expiry handling.
Once I understand it, I'll tell you what to change.

This step reveals whether Claude actually understands the code — before it starts making blind edits.

Safety Practices

12. Commit After Every Significant Operation

Don’t batch everything into one commit. If Claude does 5 things and you want to undo just the 3rd — that’s a headache.

Build the habit: Claude completes one discrete feature → commit immediately → move to the next.

Done. Write me a git commit message and I'll commit it.

13. Have Claude Write Tests, Not “Verify This for Me”

❌ "Does this code look right to you?"
✅ "Write 3 unit tests for edge cases in this function, including empty input and very large input"

Claude evaluating code it wrote itself tends toward confirmation bias. Test cases are a more objective form of verification.

14. Use /cost to Track Spending

Long session token usage can exceed expectations. Check periodically:

/cost

If costs are spiking, it usually means a task is looping through context. Find it and /clear to reset.

15. Save Effective Prompts to CLAUDE.md

When you find a prompt that works particularly well, don’t let it disappear into conversation history:

# Effective Prompt Templates

## Code Review
"Review this code as a security engineer. Focus on injection vulnerabilities and permission boundary issues."

## Refactoring
"Minimum-change refactor. Keep function signatures identical, only change internal implementation."

## Debugging
"Don't jump to a solution. List the 3 most likely causes first, and I'll confirm the direction."

Good prompts are reusable assets. Treat them like code.

A Meta-Tip

All 15 tips share a common logic: make your implicit knowledge explicit, and put it somewhere Claude can read.

The constraints you know, the pitfalls you’ve hit, the code style you prefer — all of that should live in a file, not in your head waiting to be re-explained every session.

This is the biggest difference between Claude Code and a regular AI chat box: it can accumulate context. But only if you’re willing to build that context.

Series: Claude Code in Practice · Related: Memory Architecture · Hooks in Practice · MCP in Practice