-
Setting up an AI-friendly knowledge base for your project
Your project files are the AI’s memory. If they are disorganized, the AI generates inconsistent code. I spent 3 months refining my project structure for AI readability. Here is the template. What files should every AI-friendly project have? CONTEXT.md: project overview, stack, architecture decisions. CLAUDE.md or .cursorrules: AI behavior rules and constraints. SPEC.md: active feature…
-
The art of vibe coding prompts: less instruction, more context
The best vibe coding prompts are not detailed instructions. They are context-rich descriptions. I tested this theory by comparing verbose prompts against context-rich prompts. The context-rich ones produced better code 80% of the time. What makes a prompt ‘context-rich’ versus ‘instruction-heavy’? Instruction-heavy: ‘Write a function that fetches user data from an API and displays it…
-
Iterative Prompting: Change One Thing
My first prompt never produces working code. The loop that fixes it: change one variable, regenerate, compare, keep the winner, then repeat.
-
AI Mobile Apps: React Native and Expo
The AI cannot see the screen, so mobile breaks where web does not. React Native plus Expo, tested on a real device, is the only loop that held up for me.
-
One Git Branch Per AI Session
Three AI sessions left me 47 new files and a two-hour merge conflict. Now every session gets its own branch, cut before the first prompt instead of after.
-
Debugging AI Hallucinations in Code
Fake API methods, invented config keys, library functions that were renamed. Three patterns cover almost every hallucination I have had to fix.
-
Git Workflow for AI-Generated Code
One branch per AI session, one commit per prompt that touches more than a couple of files. That habit cut my merge conflicts to almost nothing.
-
Write Specs, Not Prompts
A prompt asks the AI to guess what to build. A spec tells it. Switching to specs is the single biggest quality jump my AI output has ever had.
-
Claude Code for Backend Work
Paste an existing route above the new one and the style carries over. That is why the payment webhooks needed almost no cleanup.
-
Claude Code Plan Mode: When to Use It
I ignored Plan mode for two weeks. The first time I let it map the project before writing, the diff came back smaller and closer to right.
-
Cursor Tab vs Agent: Which to Use
Six months of doing it wrong, everything by Tab, then everything by Agent. Now under five lines is Tab, anything across files is Agent.
-
.cursorrules That Fit Your Project
Mine took eight months of tweaking. Three sections did most of the work: the stack, the architecture rules, and what I never want generated.