I Built a Website That Publishes Itself With Claude Code in 22 Minutes (Part 2)

Source: Luke Carter | https://www.youtube.com/watch?v=tyNd0wG424Y Duration: 23 min | Published: 2026-03-31 Processed: 2026-04-10

Core Concepts

A fully autonomous publishing system where an agent wakes every morning at 8am and publishes SEO-dense, GEO-optimised (Generative Engine Optimisation — ranking for AI answer engines like ChatGPT search and Google AI Overviews) articles to a blog without human intervention.
The architecture is a three-layer stack: front end (the public site), back end (the operator dashboard), and a shared Supabase database (Supabase is a Postgres database with auth, storage, and realtime built in — functions as the single source of truth).
- Front end = the domain the world sees, pulls content live from Supabase.
- Back end = "back.yourdomain.com", a login-protected dashboard for managing content, leads, emails, analytics, agents.
- Shared Supabase DB = both front and back read/write to the same tables, so a change on the back is immediately reflected on the front.
GitHub Actions runs the automation layer instead of Claude scheduled tasks or Claude Code running locally.
- GitHub Actions = GitHub's built-in scheduler/runner for scripts, baked into the repo as a YAML file, runs on GitHub's servers on a cron schedule.
- A scheduled action (e.g. "write daily article" at 8am) queries the back end for approved topics, triggers a custom API route that calls Claude Sonnet 4.6, writes the article, uploads it to Supabase, and the front end updates automatically.
- The reason for GitHub Actions over Claude scheduled tasks: Claude's scheduled tasks require the desktop app to be open continuously, which is not viable for always-on infrastructure.
Hosting and version control: codebases live on GitHub for version history, connected to Cloudflare for hosting, analytics, and A/B testing.
The operator dashboard uses a Kanban board as the content pipeline UX (columns: content ideas → approved → writing → draft/published). Replaces the Notion content calendar that was previously the source of truth.
Two modes for publishing:
- Autonomous mode — article writes and publishes directly to the live site without review.
- Safe mode — article writes to draft, author can preview, edit metadata, tweak content, and hit publish manually (human-in-the-loop).
The whole thing is distributed as two open-source GitHub repos ("Digital Home Starter" front end, "Digital Home Backend Starter" back end) plus CLAUDE.md files containing setup instructions.
- Setup flow: paste repo link into Claude Code, say "help me set this project up", Claude reads the CLAUDE.md, clones, installs deps, scaffolds templates, and walks the user through Supabase creation and API key configuration.
- Two separate Claude Code chats — one per repo — are used to avoid context pollution. The shared Supabase database links them.
Content quality is driven by a "content corpus" — a set of structured files inside Supabase that act as the brand brain:
- Core positioning, offers, competitive landscape, case studies, testimonials, SEO keyword clusters, tone of voice, banned phrases, hooks, voice guide.
- Every article generation pulls from this corpus so output stays on-brand.
The corpus is populated using two Claude Code skills (Skills = reusable named instruction bundles Claude Code loads on demand):
- Content Corpus Skill — ingests a markdown file exported from a Notion intake template (which itself is filled from an "Intelligence App" that does audience/brand research) and generates the four structured files (voice guide, core positioning, offers, testimonials).
- Content Strategy Skill — run after the back end is connected; generates topic clusters, pillar/subtopic hierarchies, and writes them into the Supabase content calendar as planned articles.
Topic clusters and semantic density are the SEO rationale:
- LLM discovery (how AI search engines surface content) rewards clustered, semantically dense content around a single subject area.
- Pillar articles link to subtopic articles within the same cluster — interlinking strategy baked into the generator.
Articles generated follow GEO best practices:
- Question-based H2/H3 titles (because users query AI in question form).
- Interlinking hyperlinks between related articles.
- FAQ blocks.
- Call to action at the end.

Buildable Ideas

Autonomous SEO publishing stack as a product offer — clone the open-source repos, white-label for clients, charge monthly for the infrastructure plus content corpus setup.
Modular agent layer on GitHub Actions — "write daily article" is agent #1, but the same pattern extends to weekly trend scans, lead enrichment, email sends via Resend, analytics reports. Each agent is one YAML file plus one API route.
Content corpus as a reusable brand-brain module — the Notion intake template + two Claude Code skills form a repeatable onboarding workflow any new client can be funnelled through in under an hour.
Kanban-driven content operations UI — the Palantir-style board treats every piece of content as a ticket moving through states, exposing review, approval, draft, and publish as explicit stages. Could be generalised into a reusable ops dashboard component.
Digital sovereignty positioning — the entire system is pitched as escape-from-GoHighLevel, escape-from-Webflow, escape-from-Notion. Jo's "own your stack" message fits this exactly. Repackageable as a Polerie offering.
Human-in-the-loop toggle as a feature — the safe-mode/autonomous-mode switch is a simple but important UX element that converts nervous clients into paying ones.
GEO optimisation layer — the article generator specifically targets AI answer engines (question headings, semantic clusters, FAQs). Could be abstracted into a standalone "GEO writer" skill that runs over any existing content library.

Key Takeaways

The architecture is reproducible in a day: front-end repo + back-end repo + shared Supabase + GitHub Actions scheduler + Claude Sonnet 4.6 for generation. No paid SaaS CMS required.
GitHub Actions is the right execution layer for scheduled agents — it runs on GitHub's servers, needs no local machine, costs nothing for public repos, and lives alongside the code it automates.
The quality of autonomous output is bottlenecked by the brand corpus, not by the model. The Content Corpus Skill + Notion intake is the real IP.
Kanban as content ops UX is a small but significant product decision — it makes the dashboard feel like mission control rather than a spreadsheet.
This is a direct template for Jo's Digital Home / Polerie play — open-source, autonomous, self-hosted, client-ready. Worth cloning the repo and studying the Supabase schema as reference architecture for Atlas.

# I Built a Website That Publishes Itself With Claude Code in 22 Minutes (Part 2) **Source:** Luke Carter | https://www.youtube.com/watch?v=tyNd0wG424Y **Duration:** 23 min | **Published:** 2026-03-31 **Processed:** 2026-04-10 --- ## Core Concepts - A fully autonomous publishing system where an agent wakes every morning at 8am and publishes SEO-dense, GEO-optimised (Generative Engine Optimisation — ranking for AI answer engines like ChatGPT search and Google AI Overviews) articles to a blog without human intervention. - The architecture is a three-layer stack: front end (the public site), back end (the operator dashboard), and a shared Supabase database (Supabase is a Postgres database with auth, storage, and realtime built in — functions as the single source of truth). - Front end = the domain the world sees, pulls content live from Supabase. - Back end = "back.yourdomain.com", a login-protected dashboard for managing content, leads, emails, analytics, agents. - Shared Supabase DB = both front and back read/write to the same tables, so a change on the back is immediately reflected on the front. - GitHub Actions runs the automation layer instead of Claude scheduled tasks or Claude Code running locally. - GitHub Actions = GitHub's built-in scheduler/runner for scripts, baked into the repo as a YAML file, runs on GitHub's servers on a cron schedule. - A scheduled action (e.g. "write daily article" at 8am) queries the back end for approved topics, triggers a custom API route that calls Claude Sonnet 4.6, writes the article, uploads it to Supabase, and the front end updates automatically. - The reason for GitHub Actions over Claude scheduled tasks: Claude's scheduled tasks require the desktop app to be open continuously, which is not viable for always-on infrastructure. - Hosting and version control: codebases live on GitHub for version history, connected to Cloudflare for hosting, analytics, and A/B testing. - The operator dashboard uses a Kanban board as the content pipeline UX (columns: content ideas → approved → writing → draft/published). Replaces the Notion content calendar that was previously the source of truth. - Two modes for publishing: - Autonomous mode — article writes and publishes directly to the live site without review. - Safe mode — article writes to draft, author can preview, edit metadata, tweak content, and hit publish manually (human-in-the-loop). - The whole thing is distributed as two open-source GitHub repos ("Digital Home Starter" front end, "Digital Home Backend Starter" back end) plus CLAUDE.md files containing setup instructions. - Setup flow: paste repo link into Claude Code, say "help me set this project up", Claude reads the CLAUDE.md, clones, installs deps, scaffolds templates, and walks the user through Supabase creation and API key configuration. - Two separate Claude Code chats — one per repo — are used to avoid context pollution. The shared Supabase database links them. - Content quality is driven by a "content corpus" — a set of structured files inside Supabase that act as the brand brain: - Core positioning, offers, competitive landscape, case studies, testimonials, SEO keyword clusters, tone of voice, banned phrases, hooks, voice guide. - Every article generation pulls from this corpus so output stays on-brand. - The corpus is populated using two Claude Code skills (Skills = reusable named instruction bundles Claude Code loads on demand): - Content Corpus Skill — ingests a markdown file exported from a Notion intake template (which itself is filled from an "Intelligence App" that does audience/brand research) and generates the four structured files (voice guide, core positioning, offers, testimonials). - Content Strategy Skill — run after the back end is connected; generates topic clusters, pillar/subtopic hierarchies, and writes them into the Supabase content calendar as planned articles. - Topic clusters and semantic density are the SEO rationale: - LLM discovery (how AI search engines surface content) rewards clustered, semantically dense content around a single subject area. - Pillar articles link to subtopic articles within the same cluster — interlinking strategy baked into the generator. - Articles generated follow GEO best practices: - Question-based H2/H3 titles (because users query AI in question form). - Interlinking hyperlinks between related articles. - FAQ blocks. - Call to action at the end. ## Buildable Ideas - Autonomous SEO publishing stack as a product offer — clone the open-source repos, white-label for clients, charge monthly for the infrastructure plus content corpus setup. - Modular agent layer on GitHub Actions — "write daily article" is agent #1, but the same pattern extends to weekly trend scans, lead enrichment, email sends via Resend, analytics reports. Each agent is one YAML file plus one API route. - Content corpus as a reusable brand-brain module — the Notion intake template + two Claude Code skills form a repeatable onboarding workflow any new client can be funnelled through in under an hour. - Kanban-driven content operations UI — the Palantir-style board treats every piece of content as a ticket moving through states, exposing review, approval, draft, and publish as explicit stages. Could be generalised into a reusable ops dashboard component. - Digital sovereignty positioning — the entire system is pitched as escape-from-GoHighLevel, escape-from-Webflow, escape-from-Notion. Jo's "own your stack" message fits this exactly. Repackageable as a Polerie offering. - Human-in-the-loop toggle as a feature — the safe-mode/autonomous-mode switch is a simple but important UX element that converts nervous clients into paying ones. - GEO optimisation layer — the article generator specifically targets AI answer engines (question headings, semantic clusters, FAQs). Could be abstracted into a standalone "GEO writer" skill that runs over any existing content library. ## Key Takeaways - The architecture is reproducible in a day: front-end repo + back-end repo + shared Supabase + GitHub Actions scheduler + Claude Sonnet 4.6 for generation. No paid SaaS CMS required. - GitHub Actions is the right execution layer for scheduled agents — it runs on GitHub's servers, needs no local machine, costs nothing for public repos, and lives alongside the code it automates. - The quality of autonomous output is bottlenecked by the brand corpus, not by the model. The Content Corpus Skill + Notion intake is the real IP. - Kanban as content ops UX is a small but significant product decision — it makes the dashboard feel like mission control rather than a spreadsheet. - This is a direct template for Jo's Digital Home / Polerie play — open-source, autonomous, self-hosted, client-ready. Worth cloning the repo and studying the Supabase schema as reference architecture for Atlas.