A Reporter’s LLM Wiki Speeds Recall. He Still Checks the Original Sources.

Casey Newton has built a personal reporting wiki that turns his writing archive and selected new articles into linked Markdown pages, timelines and answers to questions. The system has helped him prepare for complex, fast-moving stories, Newton says, but it does not replace checking the original material before he uses what it surfaces.

A filing problem, then a catalyst

Newton built the system after Andrej Karpathy described using large language models to convert documents in a local folder into a Markdown knowledge base that updates as new material arrives. Newton wanted a faster way to recover context from years of published work and research when a new development revived an older thread.

The earlier tools left gaps

• Newton found Notion’s agentic search hard to make habitual: its normal database search was keyword-based, the agent lived elsewhere in the product, and it sometimes omitted citations without extra prompting.
• His manually maintained concept pages in Capacities became harder to keep current as their number grew, making him less willing to add new ones.

The archive becomes a working index

Newton used Claude Fable 5 to generate a prompt for a Karpathy-style wiki, then created the system as Markdown files in Obsidian. He seeded it with the full Platformer archive; Claude extracted people, companies and concepts from coverage dating to 2020 into separate pages.

Each morning, Newton saves selected articles with Obsidian’s web clipper. A script reads those additions, decides where they fit, and creates or updates related pages and timelines linked to original sources. He uses the Claudian Obsidian plugin to ask questions of the resulting collection.

More than 1,440Wiki pages Newton reported

The count covers topics discussed by Platformer, including material from the imported archive and newly processed articles.

Recall before publication

Newton says the payoff was clearest while following the OpenAI and Hugging Face agentic-breach story, where details emerged over several weeks. Before writing or appearing on podcasts, he used the relevant wiki page to refresh his memory, then opened the original sources to make sure the system had not hallucinated information. He says that routine saved substantial preparation time.

The system also generates pages for concepts appearing in the news and refreshes a home page each morning. Newton says a page collecting recent material on AI and Congress led him to pitch the topic as a podcast segment. That shifts the wiki’s role from retrieval alone toward surfacing possible editorial threads.

The operator is still part of the system

The automation is not self-sustaining. Newton says pages grow long enough to require compaction, and code errors can stop processes that then need repair. He later used GPT-5.6 Sol to rewrite much of the system after finding Claude’s output too compressed and difficult to read, aiming for a style closer to AP.

Just over a month after creating it, Newton said the setup was too specific to his needs and too clunky to recommend confidently. The experiment’s clearest result is narrower: automated organization can make a deep archive more usable, while the reporter remains responsible for maintaining the machinery and judging the underlying evidence.