~/projects/ghost-blog/README.md
My long-running blog moved from Blogger to a self-hosted Ghost instance on the homelab. After the migration I ran a measured SEO campaign that restructured 15 years of content into a nine-topic taxonomy and pruned thin pages from the sitemap. On top of that sits AI-assisted publishing: a Claude Code skill that drafts and publishes posts, and a larger topic-to-article pipeline that is still being built.
cat architecture.flow
architecture.flow
Migration
SEO campaign
Publishing
cat problem.md
Fifteen years of writing lived on Blogger, a platform I didn't control. A first migration attempt never went live.
Once the blog was migrated, Search Console told the real story: 242 pages indexed, over 3,400 not indexed, and zero clicks in three months. The archive had 1,302 tags, and 1,117 of them were used by a single post. Each one was a thin page competing for crawl budget.
Writing new posts was also slow. Turning a finished piece of engineering work into a published article took as long as the work itself.
cat solution.md
I self-hosted Ghost 6 on the homelab behind Cloudflare Tunnel and wrote a converter from Blogger's Atom export to Ghost's import format. The import itself ran through Ghost's own Universal Import, driven by a signed-in browser session. That keeps every post on Ghost's supported HTML-to-editor path.
The SEO work ran as a tracked, 24-step campaign with an audit workbook as the source of truth. The tags were collapsed into a nine-topic taxonomy, keeping existing URLs where possible. Thin or duplicated posts were moved to draft rather than deleted, so every change is reversible.
Every bulk change was preceded by a full content snapshot and Ghost export, executed through the Admin API, logged step by step to a JSONL change log, and re-verified against the live site.
For new writing, a Claude Code skill turns a session into a post in my voice, generates a feature image, and publishes through the Admin API. It saves a draft by default and only publishes or schedules when explicitly told to. A fuller pipeline (topic discovery, scoring, research, generation, QC and human approval) is in progress, with its first phase complete.
ls ./components
Ghost and its database run as containers on the homelab, exposed only through Cloudflare Tunnel, with no open inbound ports.
The Blogger export is converted to Ghost's import JSON and imported through Ghost itself, so content goes through the supported HTML-to-editor conversion instead of being written straight into the database.
Day-to-day automation uses a custom integration key that can manage posts, tags and images but can't reach the database import and export endpoints.
Every URL, its topic, its decision and its status are tracked in one workbook, so the campaign is resumable and every change can be traced.
Snapshots before every batch, a JSONL change log per step, and draft instead of delete. Any decision can be rolled back.
n8n, PostgreSQL and Redis/RQ workers for topic scoring, research, generation, QC and media, with mandatory human approval before anything is published.
./run --trace
Each URL gets a topic and a decision (keep, retag, rewrite meta, or move to draft) in the audit workbook.
A full Ghost export and JSON snapshot are taken before any live change.
Changes run in batches, and each step is appended to a JSONL change log.
Updated pages, the sitemap and meta descriptions are re-fetched and checked against the intended state.
tail -f incidents.log
cat package.json | jq .stack
echo $EXIT_STATUS
outcome.log
The whole archive (1,286 Blogger posts) now lives on infrastructure I control. The sitemap is down from 1,297 posts and 1,303 tags to 632 posts and 9 topics, and 30 key posts have hand-written meta descriptions, all verified live.
Traffic results are still being measured, and a final report is due at the end of the campaign. The AI publishing pipeline is honestly marked as in progress. The skill-based publishing path is live today.