Vibe Coding: My Daily Workflow with Claude Code

This isn't a tutorial or hype piece. These are insights into how I use vibe coding efficiently in my actual work.

Before you start: Is vibe coding right for this project?

If you are using a local LLM, go ahead. If you're using commercial tools like Claude or Codex, check first, whether this project can be developed with vibe coding. It's upto client's call.

No approval? Build it yourself.
Client approves? Then you're good to proceed.

I'll break this into three scenarios:

1. A project from scratch
2. Core/major features
3. Small bugfixes

I'm assuming Claude Code for vibe coding throughout.

My standard approach:

• Skills to organize context
• @ mentions for file references
• Esc to interrupt when needed
• Plan mode for decisions
• Session management and renaming
• Hooks to protect sensitive files
• Manual commands

1. Project from Scratch

Most people say: "Use AI for planning from the start because initial setup is critical."

I disagree. Initial setup is too critical to skip.

I write my own plan first, based on the requirements. I also document the requirements it in claude.md or a separate skill depending on what I need. Then I ask Claude to create a plan in plan mode, and we compare.
I research the AI's suggestions, question them, and decide to go with either mine, theirs, or merge both.

2. Core Feature

Same process. plan debate with Claude, create a new skill for the output.
I use separate skills to save tokens and avoid hallucination.
Once we agree on the plan, I tell Claude to proceed, but I stay hands-on. I interrupt and ask why before accepting any change which I don't fully understand. If I disagree, I propose my approach instead.
This prevents sloppy code and keeps token usage down

3. Bugfixes

It depends on the bug.

• Simple bugs: I will try to identify and fix them myself. In edge cases will go with AI.
• Complex bugs: I ask Claude to find the cause and report back, not to fix. I validate the diagnosis on my end, plan a fix, then ask Claude to implement.
• If I can't figure it out: Then Claude fixes it with my manual approvals.
• Token-saving trick: I Don't paste the entire error traceback. Copy the core exception and the line number where it happened. Share that. I'll get faster, more accurate responses with half the tokens.

That's it. No magic, no shortcuts. just staying in control while letting AI handle the heavy lifting