6 Best way to use Claude Code in 2026

Over the last couple of months, like many others, I've used Claude Code a lot. I've used it to help me build projects like Build My Graphics, some other projects which are yet to launch, and lots of internal tools. I'm not using it for live coding, to be very clear about that.

I'm using Claude Code with the Max Plan, which gives you 20x the usage. You get lots of usage out of Claude Code here if you stay thoughtful and keep control. I don't loop, I don't Rolf, I stay in control.

6 Best way to use Claude Code in 2026

Stay in control

There is lots of hype about running Claude Code in a loop. The idea is that you prepare a detailed product document, then run Claude over and over again with a prompt that picks the next step and keeps going until it is done.

I totally agree that a detailed plan matters. Planning matters. But I'm not a fan of the AI working its way through a plan on its own.

I tried the Rolf loop. For me, the results were not convincing. I want to stay in control.

Plan mode

Use plan mode, because plan mode is amazing. In Claude Code you can cycle through modes, and plan mode does not instantly execute changes.

It first gathers information, explores the codebase, understands your prompt, looks up documentation if needed, and then makes a plan that you can approve or reject. It may even ask you questions if clarification is needed.

Plan mode can save bad prompts. If you're writing poor prompts, plan mode can help by asking follow-up questions to clarify missing context. That said, you should still write good prompts in the first place, because the output will only be as good as your input.

Another advantage is that it shows you what the AI wants to do, which is very valuable. It is far too easy to just trust the AI and hope it does the right thing, but you should look at its proposed steps.

Edit the plan

Do not blindly accept the plan. Edit and tweak the plan if needed.

If you asked it to fix a bug, read its diagnosis and proposed changes. If you dislike part of it, change that part in the plan and then accept.

That approach saves tokens and time. Fixing code after surprise changes is painful and often burns more tokens than editing the plan first.

Read More: desktop workflow tips for Claude Code

Agents and skills

I use agents and skills. Claude Code allows you to build custom sub-agents with their own context window to reduce pressure on the main context.

One agent I like is a Docs Explorer agent, optimized for browsing documentation. I equip that agent with tools like web search and the Context7 MCP.

I'm not a huge MCP fan because they can be token inefficient, and I prefer built-in tools. But the Context7 MCP is pretty amazing for docs. In addition to that agent, I also add project dependent skills so the AI follows my preferences.

For example, if I'm working in a Next.js project, I give it a skill file describing the best practices I want and the patterns I expect. Skills are lazily discovered, so they are only pulled when relevant, which reduces context bloat and increases the chance of a good result.

Be explicit

Explicit over implicit. The AI may sometimes do what I want, but I cannot be sure, so I am explicit.

If I am using the BetterAuth library and I want Google authentication, I state that clearly. I also instruct the Docs Explorer agent to explore the relevant BetterAuth docs before implementing.

I do not hope that it uses the right agent or reads the right docs. I tell it exactly what to do to increase the chances of the output I want.

Trust, but verify

AI is great. If you give it a good prompt with the right context, you have good chances of getting decent results.

But do not blindly trust them. Review the code carefully and think critically. Accept what is good and improve what is not.

Give the AI tools for self-verification, such as unit tests and E2E tests. It can run tests and linting commands on its own if you tell it to do so.

If you are building browser flows, tools like Playwright MCP can help, but be mindful of token cost. For step-by-step help on this approach, see automated testing with Playwright and a deeper guide on integrating Playwright via MCP.

You can still write code

You are still allowed to write code. With AI tools, it is not an either or choice.

I will not ask AI to increase a margin from 0.5rem to 1rem. I can do that myself.

If you review the code and understand the changes, getting back into the codebase is not difficult. Tokens cost money, and coding is fun, so write code when it is trivial or the AI cannot figure it out.

Setup and daily use

Install the Claude Code extension

I install the extension that works with VS Code, Cursor, and forks like Windsurf. It makes it easy to launch Claude Code inside the IDE.

I still use Cursor for Command K and tab completions sometimes, but I do most agent work in Claude Code.

I often open Claude Code in multiple panes at a time. I can run tasks in parallel when they affect different parts of the codebase in different files. If I have a file open, it is automatically pulled into context.

Launch and skip permission prompts

Claude uses a terminal UI. I was hesitant, but it is actually very good. You can tag files easily and choose what to include.

By default it prompts to edit files and run simple commands, which breaks flow. I launch it with a flag that skips permission prompts.

Run:

claude --dangerously-skip-permissions

It carries a small risk that an agent could run a destructive command, but I have never seen that happen. Choose based on your comfort level.

Switch models quickly

Slash commands are awesome. I use the model command a lot and usually prefer Opus unless Opus is having issues, then I switch to Sonnet.

Run:

/model opus

Defaults are fine for most people. It will use Opus until around half your usage then switch to Sonnet for cost. Opus is not noticeably slow and is a bit better.

Clear context often

I use clear a lot. Every time you do something new, clear.

You do not need your full chat history glued to every prompt. Compaction is another LLM call that takes time.

Run:

/clear

The up arrow returns to past chats, including prior sessions. You can close Claude, reopen later, and still jump back with ↑.

Useful terminal behaviors

Shift Enter does not add a new line out of the box. Hit Enter and ask Claude to configure the terminal so Shift Enter adds new lines. Then use Shift Enter for newlines.

You can drag files in. Hold Shift while dragging to reference them instead of opening a tab.

Pasting images from clipboard does not work with Command V. Use Ctrl+V to paste screenshots and images. That one took me a long time to figure out.

To stop Claude mid run, press Escape. Press Escape twice to browse previous messages and jump back quickly.

There is a Vim mode if you want it. I am not a Vim user, so I do not use it.

If you want to manage work in a more structured way while you queue prompts, a Kanban style approach pairs nicely with Claude. For a walkthrough, see Kanban boards with Claude.

If you are exploring image related workflows, this guide shows another path to mix images into Claude flows: image workflows with Claude Code.

Queue messages

Queuing is a feature I swear by. You can type your next prompts while the current one is running.

Queue the next task when you think of it. Claude is smart about when to run queued messages, and will pause when it needs feedback.

You can come back to a lot of work done. Still check in from time to time in case it needs your input.

Install the GitHub app for PR reviews

Claude can automatically review PRs. That is useful because using AI usually increases your PR volume.

I have seen Claude find real bugs that humans missed. To keep the noise down, customize the review prompt for bugs and security only.

Install the GitHub app via command, then edit the generated config:

claude-code-review.yaml

Use a concise prompt like this:

# claude-code-review.yaml
prompt: |
  Review this pull request for bugs and potential vulnerabilities only.
  Do not comment on style, naming, or minor nits.
  Be concise and focus on correctness, security, and unintended side effects.

Claude can also pull PR comments and address them. It can help set up your terminal behavior, too. Ask it to enable Shift Enter for new lines if needed.

Custom hooks and settings

Ask Claude to create a settings file with default hooks. This ensures it always knows about build, test, and lint without re-scanning.

Add hooks to run Prettier before edits are accepted. Add a type check after edits to ensure only correct files are accepted.

Add a Claude.md with project overview and key commands. That gives Claude a persistent reference so it does not guess build steps every time.

Create custom slash commands

You can create your own slash commands. This is great for repeatable tasks like generating tests or scaffolding a plugin.

Create a commands folder and add natural language prompts:

.claude/commands/
  test.md
  builder/
    plugin.md

Use arguments inside your prompt by inserting {{args}} in the MD file. Then call it by typing /test or /builder:plugin and pass text after the command.

This makes scaffolds and utilities trivial to trigger with one short slash command.

Memory shortcuts and hierarchical Claude.md

Use the pound sign to add memory fast. For example, type:

# always use MUI components for new UI

Claude saves that to the most relevant memory file. Claude.md files can be hierarchical.

Have one at the project root and more inside nested directories. Claude reads them all and prioritizes the most specific file when relevant.

Why Claude Code excels on large codebases

We have a React component that is 18,000 lines. There has never been an agent that could reliably update that file for us.

Claude Code has no issue updating that file. It works great with large codebases and complex tasks.

I find it gets stuck incredibly rarely. With some other tools I have to babysit more and stop jobs that were not a good fit.

Anthropic makes the best coding models and they make Claude Code the best at using the model. When they hit challenges, they improve the model and the tool together.

That focus also gives strong value for price. I pay for Max mode and I consider it a steal for the productivity it unlocks.

Practical prompts and examples

Plan mode prompt

You are in plan mode.
Goal: Fix the 404 page text and wire it to our i18n keys.
Explore the codebase, confirm the locations you will change, and propose a minimal patch plan.
Ask clarification questions if needed before proceeding.

Approve the plan only after you edit wording and scope to your liking. Edit the plan first, then execute.

Explicit docs instruction

Implement Google auth with BetterAuth in this Next.js app.
First, use the Docs Explorer agent to read the BetterAuth Google provider docs and confirm the correct setup steps for Next.js App Router.
After confirming, generate a minimal patch and unit tests.

That removes hope from the process and increases the chance the agent does exactly what you expect.

Self verification script

Run: npm run lint
Run: npm test -- --runInBand
If tests fail, fix the implementation first, not the tests.
Report failures with a short summary and the suspected root cause.

Tell Claude to run tests and lint before it proposes a final patch. Make it part of the workflow.

Terminal quick reference

Model selection:

/model opus

Start fresh:

/clear

Skip permission prompts:

claude --dangerously-skip-permissions

Stop a running task: Press Esc.

Browse previous messages: Press Esc twice.

Paste images: Press Ctrl+V.

Open past sessions: Press ↑ and scroll.

Final thoughts

Stay in control. Use plan mode, review and edit the plan, and make your intent explicit.

Equip Claude with agents, skills, tests, and hooks so it can verify itself. Trust, but verify.

Clear often, prefer Opus when you can, queue your prompts, and keep coding by hand where it makes sense. These are the 6 Best way to use Claude Code in 2026 that consistently give me strong results.