This portfolio didn't start as a weekend project. It started as a conversation with an AI agent named Arrodes.
I've been thinking about rebuilding my site for a while — the old one was stale, the structure was messy, and I kept pushing it off. So I tried something different: instead of writing every line myself, I delegated the heavy lifting to an AI assistant and focused on direction, review, and iteration.
The Setup
I'm running an open-source AI agent locally through a tool called OpenClaw. It has access to my terminal, my codebase, and my project files. I give it instructions in plain English, and it handles the rest — scaffolding components, fixing bugs, managing Git branches, pushing PRs, and even resolving merge conflicts.
Think of it like having a junior developer who never sleeps, never complains, and reads documentation at machine speed.
What It Actually Did
Here's what the agent handled end-to-end on this site:
- Scaffolded the entire Next.js project with a dark/light theme, responsive layout, and MDX blog support
- Built the contact form and wired it up to a real email API (Resend) instead of a fake
mailto:link - Created SEO infrastructure — sitemap, robots.txt, JSON-LD structured data, Open Graph tags, Twitter cards
- Refactored data files into a centralized config system so content changes in one place ripple everywhere
- Managed the Git workflow — branching, rebasing, conflict resolution, and pushing PRs to GitHub
- Switched from SMTP to Resend when I realized my free Zoho account didn't support app passwords
I reviewed every change, approved destructive commands, and made the architectural decisions. But the agent wrote the code, ran the builds, caught the errors, and fixed them before I even looked.
The Real Benefit Isn't Speed — It's Momentum
The honest truth is that AI isn't writing better code than I would. But it's removing the friction that keeps projects stuck in "someday."
When you're tired after work and staring at a blank terminal, the activation energy to start is the real blocker. With an agent, you can say "create a navbar component with these links and active section highlighting" and get something working in seconds. You review it, tweak it, and move on.
The code quality is solid enough that the real work becomes curation — deciding what to keep, what to change, and what to iterate on.
What I Learned
A few things became clear:
- Specific prompts get better results. "Make a contact form" is vague. "Create a Next.js API route that sends emails via Resend using nodemailer-style transport with loading and error states on the form" gets exactly what you need.
- Review matters. The agent made a few assumptions I caught on first pass — file structure choices, naming conventions, error handling. Catching these early is still on you.
- Local is better. Running the model locally means no API costs, no privacy concerns, and no rate limits. The tradeoff is hardware, but for a 27B–35B parameter model, a decent GPU handles it fine.
- Git is your safety net. The agent manages branches and commits, which means every change is reversible. That removes the fear of "breaking things" when experimenting.
Where This Goes
This site is a proof of concept for a bigger idea: treating AI agents as force multipliers for solo developers. The goal isn't to replace the developer — it's to remove everything between the idea and the working code.
I'll be writing more about the tools, workflows, and lessons as I push this further. This is just the beginning.
If you're curious about the stack: Next.js 16, Tailwind CSS 4, MDX, Framer Motion, Resend for email, and a locally-run Qwen model powering the agent. All open source.