For solo founders and indie hackers, the administrative burden of publishing often outweighs the creative act. Alfred Lua, a solo founder and father of two, has spent the last 75 days running a fully agentic publishing pipeline built on Claude Code. The system doesn't just store drafts; it actively researches, proofreads, tags, links, schedules, and distributes content. By treating the CMS as a set of executable skills rather than a monolithic platform, Lua has created a workflow that allows him to focus exclusively on writing.
The Stack: Obsidian, GTD, and Next.js
Luaβs architecture is a pragmatic assembly of existing tools rather than a bespoke monolith. He uses Obsidian for drafting, leveraging iCloud+ for seamless sync between devices. The heavy lifting is done by Claude Code, which interacts with two specific repositories: his website repo (hosted on Next.js and Vercel) and a Getting Things Done (GTD) repo for his content calendar. The choice of Next.js was strategic; because his articles are stored as MDX files, Claude Code can treat them as standard Markdown, bypassing the complex API integrations required by platforms like WordPress or Ghost.
Automated Research with Sonnet 5.5
The workflow begins at 6:48 am, when Claude Code automatically fetches three verified stories or quotes relevant to Luaβs themes. Using Sonnet 5.5, the agent employs web search tools to verify credibility rather than relying on potentially hallucinated training data. Lua notes that he has found this specific model version good enough for the task. Once a draft is started, the agent shifts to editor mode. Lua notes that while AI suggestions can be 'terrible,' they effectively break writerβs block by offering variations. The agent performs up to 10 rounds of proofreading per draft, catching grammar, typos, and logical inconsistencies with a speed no human editor could match.
Content Ops and the GTD Repo
Perhaps the most impressive feat is the automation of 'content ops.' Claude Code maintains a Markdown index of all previous articles, including titles, tags, and summaries. When a new draft is ready, the agent assigns three relevant tags and inserts backlinks to related past posts. It even handles scheduling logic by analyzing the content calendar stored in the GTD repo to avoid publishing similar themes on consecutive days. In one instance, the agent automatically calculated and inserted Luaβs current publishing streak into a draft, replacing a placeholder variable with the exact number of days he had been posting daily.
The Limits of AI Taste
Despite the efficiency gains, Lua acknowledges that the system is not fully autonomous. He still handles the final syndication to platforms like Substack and Medium manually, though he has automated social media posts for Buffer and Instagram. He is currently testing an 'editor' agent that learns from his revisions and successful posts to build 'lessons' and 'wins' documents. However, Lua remains skeptical about AIβs ability to replicate human taste, noting that models struggle to know when to break writing rules, a nuanced skill that remains distinctly human.
Key Takeaways
- Claude Code can effectively replace traditional CMS features like tagging, cross-linking, and scheduling when paired with a Markdown-based static site.
- Automated research agents using web search tools (like Sonnet 5.5) provide more reliable initial inputs than relying on LLM training data alone.
- While AI can handle the mechanical 'content ops' of publishing, it still lacks the tacit knowledge required for high-level editorial taste and rule-breaking.
The Bottom Line
This isn't just a productivity hack; it's a preview of the future of independent publishing. By treating code as the interface and agents as the operators, solo creators can scale their output without scaling their administrative burden.