Ten to twelve content pieces a week, across four platforms, with about two hours of human review. That's what Doneyli De Jesus, a Principal AI Architect at ClickHouse, built for his wife's solopreneur business using Claude Code and a $100/month Claude Max subscription.
The project took six months to develop and runs as a genuine production system, not a weekend demo. His wife quit corporate work to go solo, and the content treadmill was killing her momentum. A freelancer produced generic copy. Standard AI tools lacked her voice. Busy weeks meant radio silence. So De Jesus engineered something more ambitious: a three-layer autonomous system with 24 custom skills, 15 scheduled agents, and over 3,000 files.
The Three-Layer Architecture
The system runs on what De Jesus calls a "configuration-first" approach, and the layering matters.
Layer 1 is pure configuration. A global CLAUDE.md routing file (kept under 200 lines), project-specific instructions, memory files tracking metrics and strategy, and 134 files in the .claude/ directory holding shared resources. This is the foundation that gives every agent consistent context.
Layer 2 is skills: 24 structured behaviors covering drafting (LinkedIn posts, newsletters, Twitter threads), production, analytics, engagement, and research. Each skill includes its own quality gates and checklists.
Layer 3 is the agents themselves, scheduled via macOS launchd and monitored with Gatus (an uptime checker) and ntfy (self-hosted push notifications). Market intelligence agents scan signals at 6 AM weekly. Content ops agents handle planning and production. Analytics collectors and engagement monitors round it out.
The Two-Wave Monday Pipeline
Content production runs in two sequential waves every Monday starting at 1 AM.
Wave 1 uses Claude's Opus model (the expensive, most capable one) to create original content: 5-7 LinkedIn posts and one newsletter piece. Wave 2 kicks in after Wave 1 finishes and uses the Sonnet model, which is 3-4x cheaper, to adapt that content into Substack Notes and Twitter posts. The logic is sound: original creation needs the stronger model, but reformatting existing content for different platforms doesn't.
The system also pulls in external APIs for specific tasks, including Gamma for presentations and Gemini for image generation.
A critical safety detail: the system includes a human review gate. A "producer" agent checks whether the weekly content calendar was reviewed and modified by a human before any drafting begins. No approval, no content. Git commits track the planner's output and subsequent human edits, creating a verifiable chain.
Five Mistakes That Cost Months
The most useful part of De Jesus's writeup is what went wrong.
First, he built agents before configuration. The fix: work bottom-up. Get your context files and skills solid before scheduling anything autonomous.
Second, his initial CLAUDE.md ballooned to 2,000 lines. Performance degraded. The fix: keep the routing file under 300 lines and point to detailed resource files instead.
Third, voice and tone rules were duplicated across every skill. Edits required touching dozens of files. The fix: a _shared/ directory with centralized standards.
Fourth, he ran everything on Opus initially. The fix: match model to cognitive load. Opus for creation, Sonnet for adaptation.
Fifth, no monitoring. Agents failed silently. The fix: health checks and push notifications from day one, with retry logic using exponential backoff (60 seconds, then 120, then 240, maximum three attempts).
His summary principle is worth repeating: "If your best engineer can't explain the workflow in 20 bullet points, an agent can't execute it either."
Who This Is Actually For
This isn't a template you can copy-paste. Six months of development from a principal architect with 20+ years of experience is not a casual weekend project. But the architectural patterns are transferable. The configuration-first approach, the model-matching strategy for cost control, the human gate before autonomous execution - these apply to anyone building agent workflows with Claude Code, regardless of scale.
The $100/month operational cost is real, though. For a solo business producing consistent multi-platform content, that compares favorably to a freelancer or a social media manager.