After a year of experimenting with every AI tool I could get my hands on, I've settled into a workflow that actually works. Not a theoretical "here's how AI could help" workflow, but the exact tools and processes I use every day to ship code faster. Here's the full breakdown.

Stage 1: Planning and architecture (Claude)

Every project starts with a conversation with Claude. I describe what I'm building, the constraints, the tech stack, and the timeline. Claude is my go-to for architectural thinking because it handles long, nuanced conversations better than anything else. It remembers context from earlier in the chat, considers tradeoffs genuinely, and pushes back when my ideas have flaws.

My typical planning session runs about 30 minutes. I start broad ("I need a real-time notification system for 10K concurrent users") and narrow down through conversation. Claude helps me evaluate options, identify risks, and define the component boundaries before I write any code. This planning phase used to take me a full day of research and whiteboarding. Now it's one focused conversation.

Stage 2: Implementation (Cursor + Copilot)

For the actual coding, I use Cursor IDE. The inline editing (Cmd+K) handles most of the AI-assisted writing. I describe what I want, it generates it, I review and accept. For boilerplate, utility functions, and standard patterns, this is 3-4x faster than typing everything manually.

Copilot still runs in the background for tab completions. The two don't conflict as much as you'd expect. Copilot handles line-by-line suggestions while I'm typing. Cursor handles larger, instruction-driven edits. They complement each other well.

Important: I review every line of AI-generated code. I read it as if a junior developer wrote it. This isn't optional. AI code that goes unreviewed is a liability. It often works for the happy path and breaks on edge cases.

Stage 3: Debugging (Claude)

When something breaks, I go back to Claude. I paste the relevant code, the error, and my hypothesis about what's wrong. Claude's large context window means I can include multiple files without truncating. The model is excellent at tracing bugs across function calls, identifying race conditions, and spotting off-by-one errors that I've been staring at for 20 minutes.

I don't use Cursor for debugging complex issues. The inline editing paradigm works great for "change this code" but less well for "help me understand why this fails under load." Debugging needs a conversation, and Claude's chat interface is better suited for that.

Stage 4: Testing (GPT-4 + Claude)

I use GPT-4 for initial test generation because it's faster and produces concise, well-structured tests. I paste the function, describe the testing framework, and ask for comprehensive test coverage. GPT-4 usually nails the happy path tests and catches obvious edge cases.

Then I take those tests to Claude and ask for a review. "What's missing? What edge cases did these tests not cover? What would break this function that these tests wouldn't catch?" Claude consistently identifies gaps that GPT-4 missed, especially around concurrency, error propagation, and boundary conditions.

Stage 5: Code review (Claude)

Before every PR, I paste the diff into Claude and ask for a review focused on bugs, security, and performance. I explicitly tell it to skip style comments. This catches real issues about 30% of the time, things I would have missed in my own self-review. A second set of eyes, even artificial ones, is always valuable.

Stage 6: Documentation (GPT-4)

GPT-4 writes better documentation than Claude in my experience. More concise, better structured, less prone to over-explaining. I paste the code and ask for README sections, API docs, or inline comments. Then I edit for accuracy and tone.

What I don't use AI for

System design decisions on distributed systems. AI models suggest textbook solutions that don't account for your specific operational reality. Database schema design for complex domains. AI doesn't understand your business rules deeply enough. Performance-critical code paths. I write these by hand and benchmark them myself.

The cost

Claude Pro: $20/month. Cursor Pro: $20/month. ChatGPT Plus: $20/month. Total: $60/month. Sounds like a lot until you calculate the hours saved. I estimate 1-2 hours per day, conservatively. At any professional developer rate, that's a 10x return on investment.

This workflow is not about replacing my skills. It's about amplifying them. I still make all the important decisions. The AI handles the mechanical work faster than I can, freeing me to focus on the parts that actually require human judgment.