How I’m building a content agent that actually works: breakdown and architecture
@berprechtView original X ArticleView source post
CogPark
1,425 words · ~7 min read
AI Summary
The article in brief
The author describes a content operating system designed to remove repetitive production work without delegating editorial judgment. WhatsApp serves as the capture interface, Notion as the durable editorial source of truth, Drive as the media store, and modular skills as versioned workflows with explicit quality and stop rules. The agent can transcribe, retrieve references, propose connections, interview the author, and move approved assets through a scheduling pipeline. Topic choice, point of view, structure, and final approval remain human decisions. Publication approval is bound to exact text, assets, configuration, and timing, while ambiguous errors enter review rather than blind retry. The strongest boundary is that comments, DMs, and conversations remain human because they supply the lived feedback that shapes future ideas.
Suggested Lenses
Ways to explore the article
- Control BoundariesSelected
- Knowledge Architecture
- Workflow Design
Control Boundaries Deep Dive
Automation should remove operational tax, not manufacture judgment
The article separates repetitive execution from the decisions that make content distinctive. Capture, transcription, retrieval, packaging, and scheduling can be delegated because their desired outcomes can be stated and checked. Topic choice, point of view, and the connection between sources remain human because they depend on lived context and accountable judgment.
Interfaces and durable memory should be different layers
WhatsApp works well as a low-friction capture surface, but the author deliberately avoids treating it as long-term memory. Notion holds editorial state, Drive holds media, and versioned skills hold workflow rules. This division reduces the risk that useful context disappears inside chats or changes whenever the underlying model changes.
Exact approval creates a meaningful publication boundary
Approval is tied to the exact text, asset, configuration, and publishing time. Any change invalidates the prior approval, while ambiguous errors go to review instead of triggering blind retries. That turns approval from a vague permission into a reproducible control over a specific output.
Human conversation is part of the research system
The refusal to automate comments and DMs is not merely a preference for authenticity. Those conversations reveal counterexamples, user language, and weak spots in an argument. Automating them would remove the feedback loop that gives the rest of the content system useful raw material.
CogPark helps you understand the X Articles you care about with an AI Summary, suggested lenses, and a focused Deep Dive.
CogPark
Explore the next X Article in CogPark
- Open an X Article.
- Share it to CogPark.
- Read the AI Summary and explore a Deep Dive.
