Series: AI-Assisted Engineering · Part 5 of 11
- A Year of AI Coding Assistants: What I've Learned Rolling Them Out in 2024
- Model Context Protocol Explained for Engineering Leaders
- From Autocomplete to Agents: How Agentic Coding Changed Our Workflow in 2025
- AI Adoption Strategy for Engineering Teams: How to Do It Without Losing Quality
- AI Editor Wars (Cursor vs Claude Code vs Kiro): An Honest CTO's Comparison
- Spec-First AI Development in Practice: Kiro, Plan Modes and Spec Kit Compared
- Deploying an LLM Gateway with LiteLLM: Virtual Keys, Budgets and Routing Claude Code Through It
- Claude Fable 5.1 vs GPT-5.6: My Take as a CTO Who Still Ships Code
- AI Coding Accuracy and Spec-Driven Development: What Works at Startup, Scaleup and Enterprise
- Running an LLM Gateway in Your Business: Keeping Context and Code In-House (and How It Differs from Bedrock)
- When AI Writes Most of the Code, Review Becomes the Bottleneck
Contents
Disclaimer First
I'm a hands-on CTO who writes code daily. I use Anthropic's Claude heavily, so I've deliberately been careful to evaluate these tools on merit rather than habit or brand affinity. Cursor is my daily driver. Here's why, and where the others win.
Cursor: My Daily Driver
Cursor is a VS Code fork with deep AI integration. The Cmd+K inline editing, Tab autocomplete, and Agent mode (multi-file agentic editing) are genuinely excellent.
What Cursor does exceptionally well:
Tab completion that reads your mind. Cursor's autocomplete is context-aware in a way that feels different from Copilot. It understands what you're trying to do, not just what you've typed, and suggests completions that match your intent. After a year, I can't work without it.
Codebase context. Cursor indexes your entire repo and pulls in relevant context automatically, so the AI reasons about your specific code. "Refactor this function to match the pattern used in the orders module" actually works. It knows your orders module.
Agent mode for big changes. Multi-file refactors that previously took an hour now take 10 minutes in Agent mode. You describe the change, it drafts edits across multiple files, you review and approve. The review step is important: always review.
Where Cursor falls short:
The context window management can frustrate on very large repos. And when Cursor gets something wrong, which it does, the confidence with which it presents incorrect suggestions can catch you out if you're not paying attention.
Claude Code: The Terminal-Native Contender
Claude Code (Anthropic's terminal-based AI coding assistant) takes a different philosophy: it operates in your existing terminal and editor, rather than requiring you to switch to a new IDE.
What Claude Code does better than Cursor:
Agentic task execution. Ask Claude Code to "add pagination to the products API endpoint, write the tests, and update the OpenAPI spec". It will do all three steps, checking in at meaningful decision points. The agentic capability (running commands, reading files, making changes) feels more autonomous than Cursor's Composer.
Reasoning transparency. Claude Code shows its thinking more explicitly. Before making changes, it often outlines its approach. This builds trust and makes review easier.
No IDE lock-in. If your team uses mixed editors (VS Code, JetBrains, Neovim), Claude Code works for everyone.
Where Claude Code falls short:
The terminal-native workflow has a learning curve. The tab completion that makes Cursor addictive doesn't exist in the same way. For pure coding velocity on familiar code, Cursor still wins in my hands.
Kiro: The Structured Approach
Kiro (AWS's AI IDE) takes the most opinionated approach. It introduces structured concepts like "specs" and "hooks" to make AI-assisted development more repeatable.
What Kiro gets right:
The spec-driven development workflow is genuinely interesting. Write a spec (requirements, design decisions), and Kiro uses it as context for all subsequent assistance. This addresses a real problem: AI assistants often lack the why behind a feature.
Where Kiro is early:
The structured workflow has overhead that slows down rapid iteration. For mature teams with good requirements processes, it might shine. For a startup moving fast, it can feel bureaucratic. The product is still finding its footing.
My Recommendation
For individual productivity: Cursor. The tab completion alone justifies the subscription.
For agentic multi-step tasks: Claude Code. It's genuinely better at autonomous execution.
For team standardisation on structured workflows: Watch Kiro. It has a thoughtful philosophy, but it needs time to mature.
The honest CTO take: These tools are all genuinely productivity-positive. The difference between them matters less than whether your team uses any AI tooling well. Establish code review standards for AI-generated code, invest in prompt engineering, and measure the productivity impact. The tooling gap between these options is smaller than the gap between teams that use AI well and teams that don't.