I Built a Claude Code Plugin in 2 Days. Here's What Agentic Development Actually Looks Like
Forget the hype. Here's what it actually feels like to build production software with an AI agent as your dev partner, based on real projects at Lift Legal Marketing.
Everyone's talking about "agentic development" like it's some futuristic concept. Nah. I've been living it. Let me show you what it actually looks like in practice.
Two weeks ago, I built a custom Claude Code plugin for Lift Legal Marketing — a tool that lets the team interact with their entire WordPress ecosystem through natural language commands. It took 2 days. Not a prototype. A working, deployed, production-ready tool.
Here's how agentic development made that possible.
What Even Is Agentic Development?
Let me break it down simply.
Traditional AI coding = you ask for a function, you get a function. Copy, paste, move on. It's fancy autocomplete.
Agentic development = you describe a goal, and the AI reads your codebase, understands the architecture, plans a multi-step approach, writes the code across multiple files, tests it, fixes errors, and delivers a working feature.
The difference isn't incremental. It's a completely different way of building software.
Traditional: "Write a function that parses JSON" → gets function
Agentic: "Build the data pipeline for our CMS integration"
→ reads codebase → plans approach → writes 8 files
→ runs tests → fixes 2 bugs → done
The Lift Legal Plugin: A Real Example
Here's the actual project. Lift Legal Marketing manages 20+ WordPress sites for law firms. Their team needed a way to:
- Query content across all their client sites
- Generate SEO-optimized blog posts for specific practice areas
- Run bulk updates without logging into each WordPress admin panel
- Get AI-powered content suggestions based on competitor analysis
Instead of building a traditional dashboard (which would've taken weeks), I built a Claude Code plugin that wraps their WordPress REST APIs with intelligent context.
Day 1: Architecture and core API integration
I described the system to Claude Code at a high level: "We need a plugin that connects to multiple WordPress sites via REST API, with commands for content querying, bulk operations, and AI-assisted content generation."
Claude Code then:
- Read the existing project structure and package setup
- Designed the plugin architecture with proper TypeScript interfaces
- Built the WordPress API client with authentication handling
- Created the command routing system
- Added error handling and retry logic for flaky API responses
I reviewed everything, adjusted the auth flow, and refined the API client. By end of day 1, we had a working connection to all 20+ sites.
Day 2: Content generation and deployment
The AI content generation layer was the interesting part. It uses structured prompts that understand legal marketing — practice areas, jurisdiction-specific language, compliance requirements. Claude Code helped me build a prompt engineering system that produces content the Lift Legal team actually wants to publish.
By 5 PM on day 2, the plugin was deployed and the team was using it.
The Skills That Actually Matter Now
Working with agentic AI has shifted what I focus on as a developer. Here's the honest breakdown:
What matters MORE now:
- Systems thinking — You need to describe architecture, not syntax. If you can think in terms of data flow, interfaces, and component boundaries, you'll get insanely good output from AI agents.
- Critical code review — AI output is usually 85-90% right. Your job is catching that 10-15%. This requires genuine engineering knowledge. Experience matters more, not less.
- Prompt architecture — Not "prompt engineering" (that term is overused). I mean knowing how to structure a complex technical request so the AI can execute on it effectively.
What matters LESS now:
- Memorizing syntax and API signatures
- Writing boilerplate from scratch
- Manual configuration file setup
- Repetitive test writing
// My actual workflow now:
// 1. Think about the architecture (30% of time)
// 2. Describe it clearly to Claude Code (10% of time)
// 3. Review and refine the output (40% of time)
// 4. Integration testing and polish (20% of time)
Where It Breaks Down (Being Honest)
I'm not going to pretend agentic development is perfect. Here's where it still needs human brainpower:
Novel architecture decisions. When you're designing something genuinely new — not a CRUD app, not a standard dashboard, but something with unique constraints — AI defaults to familiar patterns. You need to push it in the right direction.
Domain-specific business logic. The Lift Legal plugin needed to understand legal marketing compliance rules. Claude Code doesn't know those out of the box. I had to specify every business rule clearly.
Complex multi-file refactors. When you need to change something that touches 30+ files, context windows can get strained. The solution: break it into smaller, well-defined tasks.
Debugging weird edge cases. When something fails in a way that's not obvious from the error message, you still need old-school debugging intuition. AI can help, but it can't replace the developer who says "wait, this smells like a race condition."
The Leverage Is Real
Here's what I want fellow developers to understand: agentic development doesn't replace you. It gives you leverage.
I'm a solo developer who's delivering work that used to require a small team. The Claude Code plugin for Lift Legal? A traditional agency would quote that at 2-3 weeks with a team of two. I did it in 2 days, solo.
The first version of this portfolio, game world and all? The brief said 14 days. I built it in 1. (That world was retired in July 2026 and the audit is in the repo — speed bought a codebase I could not extend, which is a lesson worth more than the number.)
That's not because I'm some genius (okay, Summa Cum Laude, but still). It's because I've learned to work with AI effectively. I know when to direct, when to review, and when to let it run.
The ceiling on what one developer can build has never been higher. If you're not exploring agentic development yet, you're competing with one hand tied behind your back.
Getting Started
If you want to try agentic development:
- Start with a real project, not a tutorial. AI agents shine when there's a real codebase to understand.
- Describe systems, not functions. Think bigger than individual code snippets.
- Review everything. Trust but verify. The AI is your partner, not your replacement.
- Iterate fast. If the first output isn't right, refine your description and try again. It's a dialogue.
The future of development isn't AI replacing developers. It's developers who use AI effectively outpacing those who don't.
I know which side I'm on.