Build a Blog From Your Terminal with Claude Code

I use Claude Code. It’s a terminal-based AI coding tool, and this blog started there - describing what I wanted built.

Terminal-first isn’t new for me. Back in 2016, writing mortgages on a DOS system, I’d run three conversations at once - one family on the phone, two more in chat - structuring loans, printing docs, running pricing. No hand leaving the keyboard.

So Claude Code felt natural. But on anything complex, I kept hitting the same wall: context management. Agents lose focus when context gets too large. They forget earlier decisions.

I needed a system - a way to structure context for agents, manage memory through the filesystem, orchestrate parallel work. So I built one, and this blog was the first test.

I’m not an engineer. I don’t write syntax. But I’m good at decomposing problems, and I’ve been practicing AI orchestration since 2022. I spend more time than I should thinking about agent memory and knowledge graphs. “What gets documented gets done” - my older brother taught me that. I wanted a blog to capture the learning.

---

The Setup

Before Claude Code touched anything, I did some manual prep:

1. Started with a quickplan.md doc in the repo - rough goals, nothing fancy
2. Had an old never-deployed domain handy (learngen.io)
3. Set up Vercel deployment connected to the repo - no Claude Code help here

After that, push to remote = blog published. That’s the workflow.

---

The Architecture

I asked Claude Code to research what features make a good blog. It came back with search, RSS, tags, reading time, table of contents, related posts. It asked my thoughts on stack - I picked from the list:

• Astro 5 + MDX for content
• Tailwind v4 for styling
• Vercel for hosting
• Cloudflare for DNS
• Dark mode because light mode hurts

Then I told Claude Code to use a parallel dev worktrees skill I wrote - a skill that creates git worktrees for each feature, spawns subagent sessions in each one, and lets them work in parallel. That’s the whole point: I designed the orchestration, Claude executed it.

(The parallel dev skill isn’t in this repo - after this project, I created a personal Claude Code PM workflow repo with even more additions on how I corral non-deterministic models with deterministic tools. But that’s for another blog post.)

Worktrees + Parallel Agents

Git worktrees let you check out multiple branches simultaneously in separate directories:

learngenblog/              # main
learngenblog-worktrees/
  ├── pagefind-search/     # feature branch
  ├── rss-feed/            # feature branch
  └── table-of-contents/   # feature branch

Each directory is independent. Run dev servers, make commits - they don’t interfere.

Claude Code spawns background agents. Each agent works in its own worktree, on its own feature:

Agent 1 → reading-time worktree → implements reading time
Agent 2 → rss-feed worktree → implements RSS
Agent 3 → table-of-contents worktree → implements TOC

When they finish, I review the commits, test locally, merge.

Worktrees combat where LLMs are weakest: context management. Each agent gets focused scope. No context bloat.

Tracking Work Against Issues

I had the agents open GitHub issues for each feature, track their work against those issues, and close them after I approved the QA.

Repo is open source - follow the work as it happens.

---

The Content

Once the app was stood up, I had a blog interviewer agent help me capture the basics. I gave it some assets to read - PDFs, Word docs, markdown files, years of unstructured notes. It created an interview I answered.

Most of what it wrote? Marked as drafts. I only published the basic info for the About page to start. I want to be deliberate about the words on this blog - the majority written by me, not AI. Review, iterate, but also: take the roast out of the oven. Put something out there for feedback. That’s why the current About is nothing too crazy.

---

LearnGen Yourself