Over the winter holidays, something shifted in developer communities. While I was scrolling through X (formerly Twitter), I noticed a pattern: experienced engineers who’d been skeptical of AI coding tools were suddenly posting screenshots of Claude Code building entire applications in under an hour. Not prototypes, but production-ready systems. So I decided to test the hype myself. I gave Claude Code a complex challenge: build a personal intelligence system that aggregates my email, calendar, notes, and creates an AI-powered yearly journal that surfaces insights I’d otherwise miss. The kind of project that would typically take days of integration work, API wrangling, and debugging.
Time to working prototype: 90 minutes
Lines of code written by me: ~30 (mostly config)
Number of times I got frustrated: Zero
Now despite the limitations, like usage caps, security issues with .claudeignore files, and costs, I understand why it’s winning compared to Cursor, Google’s Antigravity, Codex, and Replit (more on the comparison later in the post).
Before diving in, what is Claude Code?
Claude Code is Anthropic’s agentic coding tool that runs in your terminal and can act on a repository, not just chat about it: edit files, run commands, manage git workflows, and pull in external context via MCP.
What’s changed recently is not “it writes better code”, it’s that Claude Code is turning the terminal into an execution environment for intent: you describe outcomes, it plans, touches files, runs tests, opens PRs, and can keep going across sessions.
The CLI-first part matters because it makes Claude Code automation-native. IDE agents are great at tight edit loops; CLI agents are great at “do the whole thing across 12 files, run tests, fix the fallout, commit.”
The project requirements were deliberately complex to stress-test Claude Code:
Backend:
OAuth2 integration with Gmail API for email context
Google Calendar API for meeting summaries
Notion API for note aggregation
Vector database (Pinecone) for semantic search across all data sources
FastAPI server with proper async handling
Frontend:
React dashboard with real-time data streaming
Natural language query interface
Timeline visualization of insights
Export functionality for yearly reviews
Intelligence Layer:
Daily digest generation using Claude 4.5 Sonnet
Pattern detection across communications
Automatic tagging and categorization
“Year in review” generator that finds non-obvious insights
I started Claude Code with a single prompt:
Build a personal intelligence system that:
1. Connects to Gmail, Google Calendar, and Notion
2. Creates a searchable archive with semantic search
3. Generates daily AI summaries of what's important
4. Builds a yearly journal interface
5. Deploys as a web app I can run locally
Use Python for backend, React for frontend, and make it production-ready
with proper error handling and auth flows.
What happened next was remarkable:
Claude Code immediately entered “Plan Mode” (a feature from v2.1+) and outlined a 12-step execution plan covering architecture, API integrations, database schema, and deployment. Once I approved it, it worked autonomously for 74 minutes doing the following tasks:
Set up OAuth2 flows with proper token refresh logic
Built async API clients for all three services
Implemented rate limiting to respect API quotas
Created a vector embedding pipeline with chunking strategy
Generated a clean React UI with proper state management
Added comprehensive error handling and logging
Created Docker configuration for easy deployment
Wrote integration tests
The kicker: When it hit an OAuth scope permission issue with Gmail (common when integrating Google APIs), it didn’t just flag the error—it detected the issue in the error message, updated the OAuth scopes in the config, regenerated the auth URL, and continued execution.
1. Long-running autonomous execution
Claude Code can run for hours using a “compaction” strategy—when context fills up, it takes notes about where it is, clears memory, reads its notes, and continues.
2. Production-quality output
The code it generated included:
Proper environment variable management
Graceful degradation when APIs are unavailable
Rate limiting with exponential backoff
Comprehensive logging
Type hints throughout Python code
PropTypes for React components
3. Integration complexity
OAuth flows are notoriously finicky. Claude Code handled:
Token storage and refresh logic
Scope negotiations across three different APIs
Error recovery when tokens expire
Proper credential isolation
4. Natural iteration
When I asked it to “make the UI less generic—something I’d actually want to use daily,” it completely redesigned the interface with better information hierarchy, added keyboard shortcuts, and implemented a command palette for power users.
After 90 minutes, I had:
A working backend API handling three OAuth integrations
Vector search across 10,000+ emails, calendar events, notes
Daily digest generation running on cron
React dashboard with natural language queries
“Insights” feed showing patterns I’d never noticed manually
Examples of non-obvious insights it surfaced:
I have 37% more meetings on Tuesdays than any other day
My email response time correlates with calendar density (obvious in hindsight)
Three different people mentioned “investment framework” in separate contexts across email and Notion—probably worth connecting
My most productive writing happens between 2-4 PM based on Notion activity
What I had to fix manually:
Rate limiting was too aggressive (Claude played it safe)
UI color scheme needed tweaking (subjective taste)
One OAuth scope was too broad (security tightening)
Total manual code written: ~30 lines of config adjustments
Claude Code shipped v2.1.0 on January 7, 2026 with 1,096 commits. Here’s what matters:
Start coding on your laptop, teleport the session to claude.ai/code, continue on your iPad. Your full context, file edits, and conversation history move with you. This is huge for distributed work.
You can now create custom “skills”—basically specialized workflows with their own prompts, hooks, and agents—and they load instantly without restarting. I created a skill that enforces my API design patterns, and Claude started following it mid-session.
Claude Code can now control your browser directly from the terminal. When I asked it to “test the OAuth flow end-to-end,” it opened Chrome, clicked through the Google consent screen, and verified the callback worked.
Pre-tool and post-tool hooks let you intercept what Claude is about to do, more like a safety net. I set up a hook that blocks any file operations on production config files.
When you reject a plan, you can now tell Claude why you rejected it and what to change. Iteration speed went up dramatically.
I tested the same project on four alternatives. Here’s what I found:
Cursor’s strength: Better IDE integration. If you live in VS Code and want inline suggestions with minimal context switching, Cursor feels more natural.
Claude Code’s advantage: Autonomous execution at scale. Cursor excels at tactical edits—fixing bugs, refactoring functions, adding features to existing code. Claude Code excels at strategic work—building entire systems, handling multi-file dependencies, managing complex workflows.
Real user feedback (Jan 2026):
From Reddit: “Cursor excelled in backend refactoring by delivering robust modularization. Claude Code shone in integrated frontend and backend enhancements, particularly after fine-tuning prompts.”
From G2 Reviews: “Claude Code for being the most quality AI assistant for coding compared to other models available... It stands out for its exceptional ability to handle complex programming tasks, intricate algorithms, and advanced web applications.”
The elephant in the room: Google’s Antigravity launched in late 2025 offering FREE access to Opus 4.5 (the same model powering Claude Code). This should have killed Claude’s advantage.
Why Claude Code still wins:
Production stability - Antigravity is still in preview with rough edges. Claude Code is mature and reliable.
Advanced features - Session teleportation, skill system, Chrome integration aren’t in Antigravity
Terminal-native workflow - Antigravity feels bolted onto Google’s ecosystem. Claude Code was designed for terminal users from day one.
Who should use Antigravity: Developers who hit Claude’s weekly usage limits and need overflow capacity. Power users paying for both.
Different category entirely. Copilot is autocomplete on steroids. Claude Code is an autonomous agent. You don’t prompt Copilot to “build a backend API with three OAuth integrations.” You use it to complete the next line while you build the API.
Both have a place. Many developers (including me) use Copilot for tactical suggestions and Claude Code for strategic builds.
Replit Agent’s advantage: Integrated deployment. You can build and ship in the same environment with zero DevOps.
Claude Code’s advantage: Works with your existing tools, repos, and workflows. Replit locks you into their platform.
Use case fit:
Replit: Weekend projects, quick MVPs, teaching beginners
Claude Code: Professional development, complex systems, production codebases
Ethan Mollick (Wharton professor): Tested Claude Code by asking it to build a startup that makes $1000/month autonomously. It worked for 74 minutes, generated hundreds of files, and deployed a working e-commerce site selling AI prompt packs.
Casey Newton (Platformer): “Claude Code has implications for everyone, not just developers... It’s a general-purpose AI agent that can do almost anything on your computer.”
Developer testimonial (Prismic): “Bottom line: with Claude Code, I am in reviewer mode more often than coding mode, and that’s exactly how I think my experience is best used.”
Community sentiment (Jan 2026): Developers went from reviewing every step to firing multiple agents without even looking at final code. Trust increased dramatically over the holidays.
Usage Limits Are Breaking Trust
This is the biggest complaint in January 2026. Multiple sources report:
Pro plan ($20/mo): ~45 messages per 5-hour window
Max plan ($100-$200/mo): 5x-20x Pro usage, but still capped
Limits reset every 5 hours (not daily)
Weekly caps introduced in Aug 2025 after “abuse”
Real developer pain:
From The Register: “I limited out very fast this morning without even writing code—just reviewing markdown specs.”
From Medium: “Despite having an active Pro plan, I consistently hit usage limits within 10-15 minutes of using Sonnet.”
What changed: Anthropic ended a “holiday usage bonus” around Jan 1, 2026, and limits suddenly felt much tighter. Developers who’d been vibing all December hit walls in January.
The economic reality: AI inference isn’t free. Long-context coding with Opus 4.5 costs real money. Anthropic is trying to make unit economics work while keeping flat subscription pricing.
Critical bug (reported Jan 28, 2026): Claude Code ignores .claudeignore files—meaning it can read .env files with secrets even when explicitly told not to. Multiple GitHub issues confirm this.
Workaround: Use settings.json permissions instead, but syntax is confusing (absolute paths require // instead of /).
Impact: Potential for prompt injection attacks where malicious actors trick Claude into exposing credentials.
From developers:
“One day Claude gives clean code, next day it’s messy. Inconsistency makes it hard to trust.”
“Reviewing AI-generated code takes longer than writing from scratch” (for complex logic)
“Paying premium prices for something that needs heavy supervision doesn’t feel justified”
My take: These criticisms are valid for complex, novel algorithms. For integration work, CRUD apps, and standard patterns, Claude Code is genuinely faster.
Building full features - OAuth integration, API clients, dashboard UIs
Integration work - Connecting multiple APIs, handling auth flows, data pipelines
Refactoring - Large-scale code reorganization across many files
Boilerplate - Setting up projects with proper structure, tests, configs
Learning - Seeing how experienced developers structure production code
Prototyping - Weekend projects, MVPs, proof-of-concepts
Novel algorithms - Complex mathematical logic, optimization problems
Performance-critical code - Low-level systems programming, real-time systems
High-security contexts - Until .claudeignore issues are fixed
Budget constraints - Costs spiral fast for heavy users
You’re near usage limits - Frustration destroys productivity
Copilot - Tactical line-by-line suggestions while writing
Claude Code - Strategic feature builds, integrations, refactors
Manual coding - Complex logic, performance-sensitive code, security-critical paths
For professional developers: Yes, if your hourly rate justifies the cost and you understand its limitations. If you bill $150+/hour and Claude Code saves 4+ hours/month, ROI is clear.
For learning: Absolutely. Watching Claude Code build production systems teaches patterns you’d otherwise spend months discovering.
For heavy users: Frustrating. Usage limits will hit you. Consider paying for multiple tools (Claude + Cursor + Antigravity) to avoid blockages.
For startups/MVPs: Phenomenal. Speed to prototype is genuinely game-changing.
Usage limits will stabilize - Anthropic will find sustainable economics or introduce tiered “compute credits”
Security will improve - .claudeignore issues are too critical to ignore
Competition will intensify - Antigravity’s free Opus 4.5 forces pricing pressure
Hybrid workflows become standard - Multiple AI tools for different tasks
Specialization emerges - Domain-specific AI coding agents (frontend, backend, DevOps)
Claude Code in January 2026 feels like the iPhone in 2008—obviously powerful, clearly imperfect, definitely the future. The developers complaining loudest about usage limits are the ones using it most.
The real question isn’t “Is Claude Code better than Cursor or Antigravity?” It’s “What becomes possible when the bottleneck in software development shifts from writing code to specifying intent?”
My yearly journal project would have taken 2-3 days traditionally. Claude Code did it in 90 minutes. I spent that saved time making it better—adding features I wouldn’t have prioritized if I was grinding through OAuth implementations manually.
That’s the shift. AI coding tools aren’t replacing developers. They’re changing what we spend our time on. Less plumbing, more architecture. Less boilerplate, more product thinking.
Whether that future arrives in 2026 or 2027 depends on whether companies like Anthropic can solve the economics. But the technology already works.
Tried Claude Code?
I’m curious what you built and where it broke down. DM me your experience.
The views expressed are those of the author and do not necessarily reflect the views of any investment firm or portfolio company.
Comments
Nothing yet. Say the first thing.
Sign in to join the conversation.