What Works & What Doesn't (Dos and Don'ts)
✅ Strongly Recommended (Dos)
Plan Mode First Always use Shift + Tab to review the plan before letting Claude modify code. Make it a habit to immediately cancel with ESC when the plan looks wrong.
Short Cycles with TDD Maintain a short cycle of small change → lint/test → commit. This makes it easy to roll back when Claude heads in the wrong direction.
Convey Context with Images For UI tone and manner, architecture diagrams, or design references, drag and drop a screenshot instead of describing it in words. It's far more accurate than text explanations.
Paste Error Logs in Full Paste the entire terminal error log without your own interpretation. Summarizing or interpreting it actually loses information.
Be Specific with Instructions Vague instructions produce vague results. Instead of "write test code," be specific: "Write test code covering the edge case when a user logs out. No mocking."
Use think hard For complex architecture decisions or problems with many edge cases, prepend think hard to your prompt. Claude will reason more deeply. (Use only when necessary, as it increases internal reasoning tokens.)
"think hard about the edge cases in our authentication flow"Explore Before Working Before starting a large task, explicitly tell Claude to read the relevant files first and not to modify anything yet. Separating exploration from implementation prevents unintended file changes.
"먼저 인증 관련 파일들을 읽고 구조를 파악해줘. 아직 아무것도 수정하지 마."❌ Things to Absolutely Avoid (Don'ts)
Developing Multiple Features in One Session Context gets polluted and hallucinations occur. Follow the one session = one feature principle, and run /clear when a feature is done. The same applies to inserting unrelated questions mid-session.
Having Claude Directly Analyze Large Files Don't make Claude parse tens of thousands of lines of JSON or logs in the chat window. Ask Claude to write a parsing script, run it locally, and only pass the summary back to Claude.
Dumping Everything into Root CLAUDE.md The longer CLAUDE.md gets, the more Claude starts ignoring the rules. Move content like full API specs or DB schemas to separate files and split into per-folder CLAUDE.md files.
Giving Exploration Instructions Without Scope Telling Claude to "find this bug" without specifying scope will cause it to read hundreds of files and fill up the context. Always specify the search scope.
❌ "이 버그 어디서 나는지 찾아줘"
✅ "auth 모듈에서 로그인 시 발생하는 이 버그를 찾아줘"Ignoring Claude When It Repeats the Same Mistake If the same mistake keeps happening, CLAUDE.md is either too long or the instructions are too vague. Emphasize the relevant rule in CLAUDE.md with the IMPORTANT: keyword, or rewrite it more specifically.
Installing npm with sudo Don't use sudo npm install -g when you get a permission error. The correct fix is to use nvm. → Troubleshooting
💰 Checklist When Costs Spike Unexpectedly
Check these in order when you have a cost issue.
1. /context → If System Tools is large, disable MCP
2. /context → If User/Assistant is large, run /compact or /clear
3. Check CLAUDE.md length → If over 300 lines, split into per-folder files
4. Confirm you're not running multiple agents for a task that doesn't need sub-agents
5. Check if too many unrelated questions were inserted mid-conversation→ Full guide to cost optimization strategies: cost-optimization.md
💡 Decision Criteria — When to Split Sessions
If any of the following apply, run /clear and open a new session.
- You asked a question unrelated to the current task
- One feature has been completely finished
- Claude suddenly starts creating files or packages that didn't exist before
- Responses become strangely slow or Claude repeats irrelevant content
- Context is over 70% full in
/context