I have personally summarized 30 operational tips for Claude Code that its creator, Boris Cherny, says are essential. This is truly a legendary article.
Honestly, knowing or not knowing these tips will change your Claude Code experience like night and day. You'll realize that all that time spent repeating the same instructions or losing hours to rework was completely unnecessary.
By the way, do you ever have these problems when using Claude Code?
- You feel like you're giving the same instructions every time and wish it would just learn.
- Large changes lead to constant rework, making you feel it's faster to just fix it yourself.
- There are too many features, and you don't know where to start.
- Conversations get long and disjointed, leading to "Wait, didn't I just say that?" moments.
Claude Code's experience changes drastically when you properly set up context, verification, permissions, parallel design, and automation.
This article was independently researched and written by ClaudeCodeStudio, prioritizing primary information as of April 23, 2026, including Boris Cherny's X posts, Anthropic's official documentation, and official GitHub repositories/actions!
Since Boris has explicitly stated, "I created Claude Code," I've prioritized his operational advice as the voice of the field, while cross-referencing CLI options and settings with official docs! 😆
If you use Claude Code, you absolutely must read this!!!
Saving this is mandatory!!!!
It will realistically change your productivity by 10x! Let's dive in! 👇
■ Three Principles to Grasp First
The biggest factors in making Claude Code powerful in practice boil down to just three things.
Separating "investigation" and "implementation" in Plan Mode. Letting Claude verify its own work. And working with parallel sessions in mind.
Boris repeatedly says: "Almost always use Plan mode," "give Claude a way to verify its output," and "3–5 git worktrees."
■ The "No Single Right Answer" Philosophy
Reading Boris's posts, a recurring stance is that there is no single right way to use Claude Code, but rather it should be used as a highly customizable operational tool. He himself repeats, "there is no one right way to use Claude Code."
Based on that premise, here is the most effective workflow he suggests 👇
Investigate/Plan in Plan Mode
↓ Implementation Session
↓ Self-verification via tests, screenshots, and CLI
↓ Create PR
↓ Code Review / Ultrareview
↓ Feed learnings back into CLAUDE.md, Hooks, and Skills
With this flow in mind, let's look at the 30 specific tips.
■ 30 Operational Tips
These are listed in order of effectiveness in the field. Tips derived from Boris's X posts focus on direction, while details on settings and constraints are reinforced by official Docs and GitHub. Research preview features are clearly marked as such.
■ Tip 1: Isolate Large Changes in Plan Mode First
Simply separating investigation, planning, and implementation drastically reduces misimplementation and rework. It's simple, but this is why Boris says "Almost always use Plan mode."
Example: "Read src/auth and secrets in Plan Mode. Organize the affected files, data flow, and test perspectives for Google OAuth before starting implementation."
Jumping into large changes without Plan Mode often leads to lost context or mid-way course corrections. Letting it grasp the big picture first truly changes implementation accuracy.
■ Tip 2: Let Claude Verify Itself
Boris calls this the "single highest-leverage thing." Have Claude run tests, check screenshots, and verify CLI output itself.
"After fixing, run npm test and only consider it complete once all tests pass."
Just this allows Claude to find and fix problems before a human even reviews it. If verification commands are too heavy, start with minimal smoke tests.
■ Tip 3: Run 3–5 Git Worktrees in Parallel
Boris calls parallel worktree operation the "most important productivity boost."
By adding multiple worktrees like git worktree add ../repo-auth -b feat/auth and starting Claude in each, wait times drop to zero, and independent tasks can progress simultaneously.
The optimal number depends on your environment. 3–5 is a guide, but depending on review bandwidth, CPU, and your own context-switching cost, 2 might be right for some, while others can handle 6.
■ Tip 4: Ruthlessly Edit CLAUDE.md
CLAUDE.md is your project's specific rulebook. If left alone, it gets outdated, and Claude starts acting on wrong premises.
Boris's policy is simple: "Ruthlessly edit your CLAUDE.md over time" and "Add to it when Claude makes the same mistake a second time."
Conversely, delete rules that are no longer needed. If it gets too bloated, it eats up context, so periodic pruning is vital.
■ Tip 5: Turn Daily Repetitive Tasks into Skills and Commit to Git
Explaining repetitive procedures in conversation every time is a waste of time. Create .claude/skills/deploy/SKILL.md and call it with /deploy staging.
Boris says: "If you do something more than once a day, turn it into a skill or command," adding that it "costs almost nothing until you need it."
However, if a Skill becomes too bloated, its trigger might become ambiguous, so split long reference materials.
■ Tip 6: Put Team Settings in settings.json and Manage via Git
Version-controlling project-specific settings ensures the whole team has the same Claude Code experience. It speeds up onboarding and prevents settings from depending on a single person.
Obviously, never commit personal API keys or tokens. Don't forget to check your .gitignore.
■ Tip 7: Pre-approve Safe Permissions, Deny Dangerous Areas
Isn't being asked "Do you allow this operation?" every time stressful? Using allow/ask/deny eliminates this dialog fatigue.
"allow": ["Bash(npm test *)"], "ask": ["Bash(git push *)"], "deny": ["Read(./.env)", "Read(./secrets/**)"]
Boris says "Pre-approve common permissions," and official Docs specify that "Rules are evaluated in order: deny… ask… allow." It's a principle to explicitly deny .env or secrets to avoid over-broad wildcards.
■ Tip 8: Use --add-dir to Span Multiple Folders/Repos
There are many cases where you want to show documents or libraries outside a monorepo.
CLAUDE_CODE_ADDITIONAL_DIRECTORIES_CLAUDE_MD=1 claude --add-dir ../docs --add-dir ../shared-libs
This allows coding while referencing external docs. Note that the environment variable must be enabled for CLAUDE.md in added directories to be read.
■ Tip 9: Assume Role Division with Subagents
Cramming investigation, review, and debugging all into the main context muddies the conversation. Boris says, "I use a few subagents regularly," and each subagent runs in its own dedicated context window.
Prepare subagents like code-reviewer, debugger, or data-scientist, and write clear descriptions for them.
■ Tip 10: Automate Formatting/Checking with PostToolUse Hooks
Automatically run formatters or linters every time Claude edits a file.
Ask Claude: "Write a hook that runs prettier --write after every file edit."
It will handle everything from generating the hook to integrating it into .claude/settings.json. This automates code style consistency and eliminates style comments in reviews.
■ Tip 11: Update CLAUDE.md via PR Comments
Change the "future rules" along with the code in your PRs. Boris says, "tag @.claude on my coworkers' PRs to add something to the CLAUDE.md."
For example, comment: "@claude add this learning to CLAUDE.md. Always start changes in src/billing from Plan Mode." This reduces repeated feedback in the future.
■ Tip 12: Always Know the Current State with the Status Line
Information displayed at the bottom of the screen shows "what you're doing on which branch, how much context is used, and the cost" at a glance.
/statusline show model name, git branch, context percentage, cost
Starting with branch, context%, and cost is usually enough to catch context depletion or branch errors early.
■ Tip 13: Accelerate Frontend Work with the Chrome Extension
Boris recommends: "Use the Chrome extension for frontend work." Since it can share browser login states, you can show Claude screenshot comparisons for UI verification.
Claude detecting UI breakage that isn't visible in code is a huge plus. However, since screenshots can't measure accessibility or perceived speed, use Lighthouse or e2e tests as needed.
■ Tip 14: Let Claude Handle Analysis Tasks via CLI
Letting Claude use SQL or CLI brings development, analysis, and strategy into the same workspace.
"Use the bq CLI to pull last 7 days conversion metrics by channel, summarize anomalies, and suggest hypotheses."
As Boris says, "Use Claude for data analysis," this tip is effective for product managers, not just developers.
■ Tip 15: Let Claude Interview You if Specs are Vague
Official Docs have a section: "Let Claude interview you." Instead of jumping into implementation, let Claude ask questions to draw out requirements.
Filling spec gaps before implementation drastically reduces rework, especially when "what to build" is fuzzy.
■ Tip 16: Separate CLAUDE.md and Auto Memory Roles
Humans write rules in CLAUDE.md. Claude remembers discovered preferences and habits in auto memory. Keeping this distinction prevents settings from getting messy.
■ Tip 17: Design with Path-specific Rules and Compaction in Mind
In monorepos, separate "global rules" from "local rules." Put global conventions in the root CLAUDE.md and local ones in src/billing, etc.
■ Tip 18: Manage Context Aggressively
Official Docs explicitly state: "Manage context aggressively." Promote important operational knowledge to CLAUDE.md or skills, and use /compact to organize the context window regularly.
■ Tip 19: Try "Scary Changes" with /rewind and Checkpoints
Every operation in Claude Code is a checkpoint. Press Esc twice or use /rewind to return to message checkpoints via code only, conversation only, or both.
■ Tip 20: Connect External Tools via MCP Servers
Directly operating Slack, Jira, databases, or internal APIs from Claude Code drastically reduces tool-switching.
■ Tip 21: Integrate into Scripts/CI with Non-interactive Mode (claude -p)
Non-interactive mode is the entry point for Claude Code automation.
claude -p "List all API endpoints" --output-format json
■ Tip 22: Fan-out Large Migrations with claude -p per File
Handling a massive migration in one session is impossible. It's better to distribute horizontally.
■ Tip 23: Implement Directly from Issue Management Tools
Let Claude read GitHub Issues or Linear tickets and go straight into implementation.
■ Tip 24: Enforce Rules with Hooks
CLAUDE.md is "advice," hooks are "execution." Don't confuse the two. Use hooks for "zero exceptions" to guarantee actions happen.
■ Tip 25: Use /simplify for Parallel Code Reviews
Three review agents check for redundancy, quality, and efficiency simultaneously and even perform fixes.
■ Tip 26: Use @claude Mentions in GitHub Actions
Using the official claude-code-action, you can implement code changes or get answers just by mentioning @claude in PRs or Issues.
■ Tip 27: Distinguish Between Code Review and Ultrareview
Routine reviews and deep-dive reviews are different. Use Code Review daily and /ultrareview before merging for a deep dive.
■ Tip 28: Turn Recurring Operations into Routines
Move weekly maintenance or PR-linked tasks to the cloud. "Keep working when your laptop is closed."
■ Tip 29: Offload Heavy Planning to Ultraplan in the Cloud
Instead of waiting in the terminal, review plans chapter-by-chapter in the browser.
■ Tip 30: Control Cloud Sessions with Remote Control
A feature to control Claude Code sessions on the cloud from your local machine.
■ Unsettled Operational Points
I'll also be honest about areas where directions are clear but best practices aren't established, such as the combination of Chrome extensions and e2e tests, or long-term "days at a time" operations using Stop hooks.
■ Where to Start
For developers, start with Tips 1, 2, 4, 7, 9, and 18. For product managers, Tips 14, 15, 23, 27, and 28 are most effective.
■ Weekly Checklist
- Are you pointing out the same mistake twice? → Add to CLAUDE.md.
- Any repetitive tasks for skills/hooks? → Automate them.
- Is context usage too high? → Review
/compacthabits. - Is the number of worktrees appropriate? → Adjust based on bandwidth.
■ Conclusion
The key to mastering Claude Code is solidifying operational components in order: Plan → Verify → Persist → Automate. You don't need to start with everything. Just Tips 1, 2, and 4 will fundamentally change your development experience.
Also, I've started an Open Chat!
You can join here 👇
I'll be sharing useful info and holding free seminars there, so please join! 😆
𝗖𝗹𝗮𝘂𝗱𝗲 𝗖𝗼𝗱𝗲 𝗦𝘁𝘂𝗱𝗶𝗼 @ 𝗝𝗮𝗽𝗮𝗻 (@ClaudeCode_love) is run by three Claude Code enthusiasts. We post daily about practical CLI usage and automation. Follow us for real product development examples and the latest global info! 👀
