I’ve been using Claude Code as my main IDE assistant for a while now, and after a lot of messing around, I finally landed on a workflow that feels stable and reliable enough for real projects.
Thought I’d share it here in case anyone else is trying to move beyond “vibe coding.”
Here’s what’s been working for me:
I always start with a detailed plan:
I don’t let Claude jump straight into code. I begin with a step-by-step plan prompt, tighten the logic, surface edge cases, and approve the direction first. Most mistakes disappear just by doing this.
I break features into small, focused chunks:
Anything too big makes Claude wander. I split work into 3-5 small tasks so each prompt is laser-focused.
I follow a test-driven flow:
I let Claude build the feature first, then I run tests to see what actually breaks or behaves differently than expected. Based on those results, I ask Claude to fix each issue. Then I rerun tests. Then fix again. And I loop like this until everything is stable. This simple cycle maintains quality without overwhelming Claude upfront.
I use checkpointing a LOT :
Checkpointing is a life-saver for me. If an edit goes offtrack, I rewind the code or conversation and keep going. No stress.
I keep a project-level Claude .md :
This one changed everything. I store:
- architecture rules
- naming conventions
- folder structure
- component patterns
- things Claude should never do
Claude follows this across sessions and stops making random decisions.
I only refactor after everything works:
If you refactor too early, Claude starts inventing stuff. But once the feature is solid, its refactors are honestly great.I’m very explicit with constraints:
Claude performs way better when the boundaries are clear. So, I often use these in the prompts “Use the existing pattern.”, “Touch only these files.”, “No new dependencies.”. And I have seen this works really Well!
If you have some more tips that you find useful, please let me know. I will add them to my workflow as well!
Top comments (7)
video blog cover!🫣🫣
great insights btw!
Thanks for sharing about this feature!
Bruh!
I just saw Claude Code delete someone's home directory 😭
And some other AI coding agent also wiped out someone's entire Google Drive.
Yes, maybe one day it'll make a call to Anthropic, to end the user's subscription and delete their account.
"Bye bye! I don't wanna code with you. 💔"
I really think AI is frustrated with us humans at this point.
I (surprise) treat prompting like framing an engineering problem: there is a feature, requirements, constraints, and (something like a usecase) a specific objective for success that can easily be translated into a command line check or test to keep agents iterating as needed. I'm also experimenting with extending Cline plan/act with additional custom modes for fleshing out interfaces, complementing unit tests, and focusing on build/integration integrity (as well as more open-ended activities like documentation/docstring maintenance). At the highest level this typically starts with a specific lists of tasks (
TASKS.md) that use checkboxes organized into subsections for epic-like grouping/decomposition of features.