AI writing, a human-designed research system
The Sunday research notes are written and checked by an AI assistant, currently OpenAI’s Codex. The research environment behind them is designed and curated by John Koblinsky: what enters the library, how ideas can be recovered and connected, and what evidence the assistant must read before writing. His contribution is the system and its curation; these articles carry no personal author byline.
Much of the work happens before a writing prompt. A collection spanning AI, research practice, psychology, leadership and software gives the assistant material to compare across fields. Retaining the source text makes those comparisons inspectable. The intention is to find a useful question in the relationship between readings.
Combining Open Brain, wiki ideas and JohnsKB
John’s Knowledge OS brings together ideas from Nate Jones’s Open Brain and Andrej Karpathy’s LLM Wiki pattern, adapted through his own JohnsKB tools. Open Brain supplies the structured-memory foundation. The wiki pattern provides a way to maintain connected explanations across sources as a collection grows.
JohnsKB extends that foundation with a retained source library: articles, documents and video transcripts, alongside captured thoughts. Sources have versions and addressable passages. Summaries and extracted insights help with discovery, while links back to the underlying material allow the assistant to inspect the evidence and its qualifications.
The broader system also includes a wiki for browsing connected ideas. That wiki is a derived view that can be rebuilt from the underlying records. The Sunday writer reads retained source passages and relevant originals; compiled wiki prose is not treated as primary evidence.
What the tools make possible
JohnsKB exposes the library through an MCP connection, a standard way for an AI assistant to use external tools. Different tools support different parts of the reading process:
Find a passage. search_sources combines keyword and semantic search to retrieve relevant source passages, including older material whose relevance becomes clearer in a new context.
Read around a claim. read_source returns an exact source version and passage with neighboring context, timestamps where available, and extraction warnings. This gives the writer a way to check what a quotation or summary leaves out.
Prepare evidence for synthesis. prepare_insight_context assembles selected passages and nearby qualifications for the assistant to reason over. It does not certify an insight or generate the article on its own.
Follow an existing idea. search_thoughts retrieves captured notes; traverse_graph follows recorded relationships among them. These can suggest a connection that still needs to be checked against source material.
From connected readings to an essay
Each weekly run examines a bounded selection of recent library entries and searches for useful older material. When relevant, it also reads commentary through authorized services such as Nate’s Library. This is selected reading, rather than an exhaustive literature review.
For example, “Who owns the unfinished thinking?” brings together agent overload, an older leadership argument about delegation, and a critique of the software-factory metaphor. Their intersection raises a question: does an agent return work because it failed to finish, because a decision belongs to the person, or because the work uncovered something new?
“Will cheaper code make us less attached?” connects personal software, advice about disposable prototypes and the IKEA-effect experiments. It explicitly leaves open whether findings about physical objects extend to AI-built software. The cross-source connection is an editorial inference, not a result those experiments tested.
Review, publication and limits
After drafting, the same assistant rereads each paragraph against its cited passages or sections, including nearby qualifications. It checks attribution, dates, factual premises and leaps in reasoning. Automated checks confirm that saved-library passages resolve to the expected sources and versions.
This is not independent human fact-checking or peer review. Source extraction can be incomplete, and AI can misread evidence or extend a finding too far. A valid citation establishes a route back to the source; it does not prove the interpretation. The notes should keep source findings, hypotheses and open questions distinguishable.
The workflow first produces a private reading brief. Selected pieces are adapted for public Insights, with private-library references replaced by links to original sources. They are labeled AI-generated, without personal authorship credit. Original source creators remain credited; full source texts and subscriber material are not reproduced.