Claude Code was HARD Until I Learned These 6 Things
How I managed to use Claude Code to do real work
Having a hard time getting Claude Code to do anything more than build simple web apps for a demo?
You're not alone.
I've spent over 10 years building software, from trading systems to data pipelines to my current role as an AI Research Engineer. And I’ve struggled to find the right balance between AI code tools and my human brain.
After a month of focused practice, I've reached a much better place and want to share the six rules that transformed my approach.
Rule 1: Start with the End in Mind
Define what success looks like before you code.
Initially, I gave these tools too much credit and skipped the planning phase. Now I spend upfront time thinking through what I actually want to build.
Key actions:
Use AI to help with planning, but invest time in the process
Create a markdown file capturing the context and goals
Take this document with you to your coding sessions
Rule 2: Build Validation Cycles Upfront
Spend more time developing constraints than implementing code.
With AI code tools, the biggest shift is that more time goes to developing constraints rather than writing the actual implementation.
The mindset shift:
Focus on setting up testing and validation frameworks first
Define how you'll prove the code works
Slowing down here makes the coding part instantaneous
Rule 3: Don't Let AI Run Too Far Ahead
Constrain the model to work step by step.
The more the AI does in one shot, the more you'll race to keep up. This leads to:
Lost context
Inability to redirect when things go wrong
Subtle bugs slipping through
Best practice: Always constrain it with "implement this step by step pausing to test each milestone" rather than letting it build everything at once.
Rule 4: Use Multiple AI Windows
Separate planning, implementing, and learning into different chat threads.
My three-window setup:
Planning agent: Talk through project context, brainstorm improvements
Implementing terminal: Write code step-by-step within well-defined constraints
Explainer chat: Ask questions about the generated code without cluttering the main thread
This prevents the main coding thread from getting clogged with tangential discussions.
Rule 5: Remember Software Engineering Fundamentals
AI tools amplify your good and bad qualities as an engineer.
Just like winning the lottery doesn't change who you are but amplifies it, AI coding tools make your engineering strengths and weaknesses more visible.
Essential principles to explicitly call out in prompts:
Write functions that do one thing
Build composable, extendable components
Design for easy testing
Follow clean code principles
Make systems modular with functions grouped by purpose
Rule 6: Know When to Clear Context
Recognize when you're going in circles and start fresh.
When the AI keeps saying "you're absolutely right, let me do it this way instead" and you're getting frustrated, it's time to reset.
How to reset effectively:
Ask the model to dump the current conversation state
Write a to-do doc capturing where you left off
Start a new chat thread with that context
Take a break yourself
The Bigger Picture
We're all figuring this out together. These tools are constantly evolving, and you're not behind—you're still early.
My advice: Rather than trying to keep up with everything, pick one tool and go deeper than you normally would.
Learn something that will be useful tomorrow.
This is exciting but overwhelming territory. Focus on sustainable progress over breadth-first exploration.