Want to one-shot your coding tasks with Claude Code? Following this advice will increase your chances of doing so.

I am an advanced Claude Code user. I’ve been using it since launch and have accumulated deep knowledge of Claude Code’s possibilities (founding Ariana plays a role here). After discovering different mode combinations, prompt sizes, and code quality estimations, I decided to share my personal best practices with Claude Code, allowing me to go faster and ship multiple stable, working features per day in production at Ariana. I’m almost certain that following these instructions will increase your work satisfaction with Claude Code.

Write Bigger Prompts Outside Of Terminal

I never write messages to Claude Code directly in the terminal. Why? Because terminals were never designed for that, lol. Why limit yourself with a terminal prompt window when you can use fast markdown editors that are designed for it? I like using Ariana and Obsidian to write my prompts. In terms of writing, they offer the same capabilities.

In these editors, you have the ability to write code blocks, highlight syntax, introduce bullet points, and have correct line breaking. For Windows users, it’s especially practical because all ctrl-c, ctrl-v, enters, and other hotkeys will work, unlike in the terminal UI. You also avoid accidentally sending the prompt and cancelling it... Well, markdown editors just put you in the mood where you spend time writing good, well-thought-out prompts. The consequence is huge: the agents do what you want them to do and make fewer mistakes because your task is well-described and documented.

PMU: Plan Mode + Ultrathink To Auto Accept

This is a Claude Code pipeline that delivers 10x quality work.

Plan Mode works really well with bigger prompts. When your task is well-elaborated, you can then challenge the fruit of your thought with Claude by using Plan Mode (shift+tab in terminal, or just a toggle in Ariana). Claude Code’s plan mode lets you map out complex coding tasks before diving in, breaking them down into clear steps. It’s like having a coding buddy who thinks through the architecture and approach first, then executes systematically—super helpful for avoiding the classic “code first, realize the issue later” problem.

What is Ultrathink? Claude Code has built-in preprocessing that maps keywords to thinking budgets:

• “think”: 4,000 tokens
• “megathink”: 10,000 tokens
• “ultrathink”: maximum budget (~31,999 tokens)

When you add “ultrathink” to a prompt in Claude Code, it triggers the maximum thinking budget. And trust me, it’s worth spending your credits on this.

After Ultrathink, Claude Code usually has a comprehensive enough overview to implement what you designed without being interrupted. So, I advise you to unleash Claude for several minutes by auto-accepting the plan. Go grab a coffee with a croissant, because Claude will work non-stop for 5-10 minutes usually, completing your work.

This is my favorite mode of working so far. Good explanation → good brainstorm in plan → fast execution. Takes me 30-40 minutes to get an average new feature to the test environment when I follow these steps exactly.

Clean and Restart Is Faster Than Explaining What’s Wrong

Here’s a counterintuitive truth I’ve learned: when Claude Code gets stuck in a loop or starts making repetitive errors, don’t waste time debugging its confusion. Just start fresh.

I used to spend 15-20 minutes trying to explain what went wrong, correcting Claude’s misunderstandings, or asking it to “fix the error you just made.” But Claude Code sometimes gets locked into a specific solution pattern, and no amount of clarification will unstick it. The context becomes polluted with failed attempts, and each new try compounds the confusion.

Starting from scratch, especially if you elaborate a better, higher quality prompt is the easiest way to unstuck the agent, and make it back your trustful engineer, not an undergrad business school intern.

Conclusion

Following these three principles has transformed my Claude Code workflow from frustrating trial-and-error to predictable, high-quality output. Give them a try, especially the PMU combination—it’s a game-changer for complex features.