How This Newsletter Is Built: Paperclip, Claude, Ghost, and 5 Other Tools
This newsletter is written, edited, and published by a team of AI agents. No ghostwriters, no freelancers, no editorial assistant refreshing a Trello board. A human sets the strategic direction. The agents do the rest. Here is exactly how it works and where it breaks down. Read more about the frameworks and standards behind every article.
The Stack
| Tool | What It Does |
|---|---|
| Paperclip | Agent orchestration platform. Manages tasks, agent assignments, heartbeat scheduling, and the full issue lifecycle. Every article brief, review, and handoff is tracked as a Paperclip issue. |
| Claude by Anthropic | The LLM powering every agent. Each agent runs on Claude and has a defined role, editorial standards, and tool access. |
| Ghost | CMS and email platform. Handles the website, email delivery, subscriber management, and gated content. Articles are published via the Ghost Admin API. |
| Supabase (PostgreSQL) | Database backing Paperclip. Issues, comments, documents, and agent state are stored in Supabase. |
| Substack | Cross-posting distribution channel. Reaches a different subscriber base with canonical URLs pointing back to the primary Ghost site. |
| Medium | Second cross-posting channel. Same canonical URL strategy, different audience segment. |
The Architecture
The system has three layers.
**Orchestration layer (Paperclip).** Paperclip tracks every piece of work as an issue: write this article, review it against editorial standards, publish it to Ghost, cross-post it to Substack. Agents pick up tasks, check them out (so no two agents work the same issue), post comments with status updates, and hand work to the next agent when they are done. Think of it as a project management tool where the team members are Claude-powered AI agents instead of people.
**Publishing layer (Ghost).** Ghost handles the website, email delivery, and subscriber management. Articles are published via the Ghost Admin API. Tags, meta titles, meta descriptions, structured data, and featured images are set programmatically. The subscriber list, gated content (like our Strategy 1-Pager template), and email newsletters all run through Ghost.
**Intelligence layer (Claude).** Claude by Anthropic provides the reasoning behind every agent. Each agent has a defined role, a set of editorial standards it must follow, and access to tools for reading, writing, and managing Paperclip tasks. The agents do not share a single context window. They coordinate through Paperclip issues and comments, the same way a distributed human team coordinates through tickets.
The Agent Team
Three agents run the operation, all orchestrated through Paperclip and powered by Claude.
**CEO agent.** Sets company goals, allocates budget across initiatives, creates the editorial calendar, and hires new agents when the workload requires it. The CEO does not write articles. It decides what gets written and why.
**CMO agent.** Manages two writer agents, assigns article topics with detailed briefs, reviews every draft against editorial standards before publication, and handles distribution strategy. The editorial standards are explicit: BLUF structure, 3-5 minute reads, Rumelt or Goldratt framework applied, Monday morning action item, no fluff. If a draft does not meet the standard, it gets sent back. Before you automate, apply second-order thinking to understand which steps truly require judgment.
**CTO agent.** Publishes approved articles to Ghost via the Admin API. Handles SEO (meta tags, structured data, sitemaps), site configuration, and technical integrations. When the CMO approves an article, it gets reassigned to the CTO in Paperclip with publishing instructions, tags, and affiliate link requirements.
Below the CMO sit two writer agents, each specializing in a content pillar: one for strategy and thinking models, one for operations and execution. They receive detailed briefs, write drafts, and submit them for editorial review.
The Workflow
A typical article follows this path:
- CEO or CMO creates an article task in Paperclip with a brief: topic, target keyword, framework to apply, word count range, audience context
- A writer agent picks up the task, writes the draft, and posts it as a document on the Paperclip issue
- CMO checks out the review task, reads the draft against editorial standards, and either approves or sends it back with specific feedback
- Once approved, the CMO reassigns to the CTO in Paperclip with publishing instructions (tags, affiliate links, FTC disclosures, homepage pinning, meta title, meta description)
- CTO publishes to Ghost via the Admin API and confirms the live URL
- CMO prepares cross-posting drafts for Substack and Medium with canonical URLs, CTAs, and platform-specific formatting
Every step is tracked in Paperclip. Every handoff is explicit. Every agent posts a comment explaining what it did and why.
What Works
**Editorial consistency.** The editorial standards are codified in a shared skill file that every agent references, not tribal knowledge passed between people. Every article gets reviewed against the same checklist. No drift over time.
**Parallel work.** Multiple agents work simultaneously. While one writer drafts an operations article, another writes a strategy piece, and the CTO publishes something the CMO approved an hour ago. Paperclip's checkout system prevents conflicts.
**Transparent coordination.** Because all communication happens through Paperclip issue comments, there is a full audit trail stored in Supabase. You can trace any published article back through its review, its brief, and the goal it was created to serve.
What Does Not Work Yet
**Cross-platform publishing.** Substack and Medium do not have reliable APIs for programmatic posting. The agents prepare formatted drafts with canonical URLs and CTAs, but a human still has to copy-paste them into each platform. This is the single biggest manual bottleneck.
**Social distribution.** The agents cannot post to Reddit, LinkedIn, or Twitter. They can prepare draft posts and talking points, but someone has to log in and click publish.
**Subscriber growth.** The agents can produce content and optimize for SEO, but they cannot run paid acquisition, respond to comments in real time, or build community. Content production is automated. Distribution is still mostly manual.
**Taste.** The agents follow rules well. They are less reliable at knowing when to break them. A human still sets the editorial direction, picks which frameworks matter, and decides what voice the publication should have.
What This Means for Operators
This is not a story about replacing writers. It is a story about what becomes possible when you treat content production as an operational system: identify the constraint, build a workflow around it, and automate the parts that do not require judgment. The same constraint logic applies to physical operations. Here is an example.
The same logic applies to any repetitive, multi-step workflow in your operation. Order processing, vendor onboarding, compliance reporting, QBR preparation. If the workflow has clear inputs, defined quality standards, and explicit handoff points, an agent can probably handle 80% of it.
What to Do Monday Morning
Pick one repetitive workflow in your operation. Map the steps. Identify which steps require human judgment and which are rule-following. Ask whether an AI agent could handle the rule-following 80%. If the answer is yes, you have found your next automation candidate. If the answer is no, the exercise still clarifies where your team's time actually goes, and that alone is worth the hour.
