consulting-writing

LLM Wiki Newsroom

A multi-agent AI knowledge base run by a five-role "newsroom" — open-source, local-first, no vendor lock-in. Drop articles, documents, and PDFs into the raw/ folder, type a single command, and the newsroom — powered by an agent like Claude Code — reads them, extracts entities, concepts, and relationships, and organizes everything into a fully cross-referenced wiki, a structured and persistent alternative to RAG. Unlike most takes on the idea, the agent that writes a page is never the one that reviews it, and the authoring guidelines evolve themselves over time. Every new document you add also enriches the existing pages. This repo ships with a small example corpus — the debate over what "open source" means for AI — under wiki/, but the framework is domain-agnostic.

Most knowledge tools leave the finding to you. This project makes the AI read and understand your collected documents first, then organizes them into a wiki — with cross-references between pages, automatic flagging of conflicting claims, and per-topic synthesis built in from the start, so later retrieval is fast.

See the output before installing — the example corpus shipped in this repo is published as a browsable GitHub Wiki (no clone needed). It's a rendered static snapshot of the wiki/ folder; the interactive graph below runs locally.

The interactive knowledge graph (graph/graph.html) — every page a node, every wikilink an edge, color-coded by auto-detected cluster, with a live physics layout and filter/search built in. Shown here on a larger private deployment (~2,300 nodes) to convey how it scales; this repo ships a deliberately small 15-node example corpus you can browse the exact same way. (Interface shown in the optional Korean WIKI_LANG=ko mode.)

The concept

This project is one question worked into a running system: how far can you trust knowledge an AI wrote? Two ideas organize everything below.

1. The product is an LLM Wiki — Andrej Karpathy's three-layer pattern. The original documents you collect (Layer 1, raw/), the cross-linked wiki the agent maintains (Layer 2, wiki/), and the operating rules the agent follows (Layer 3, CLAUDE.md + .claude/) are kept strictly separate, so humans and AI don't trespass into each other's territory. As the operator you do exactly two things — feed Layer 1 and tune Layer 3; only the agent writes Layer 2. And ingesting one document doesn't just add a page: it refreshes the ~10–15 existing pages that mention the same entities and concepts, which is what makes the wiki compound instead of just piling up.

2. The factory is a newsroom running four loops. The wiki is produced by five roles modeled on a newspaper staff — and the agent that writes a page is never the one that reviews it:

Four loops turn that division of labor into trust. The first three nest inside one another; only the fourth sits outside, feeding published pages back in:

%%{init: {"flowchart": {"wrappingWidth": 700}}}%%
flowchart BT
    subgraph meta["Meta loop — mistakes become new rules"]
        subgraph outer["Outer loop — two gates before publication"]
            inner["Inner loop — the writer self-checks while drafting"]
        end
    end
    %% the   padding keeps this box as wide as the Meta loop box above
    reground["  Reground loop — pages that go stale or inconsistent come back as input  "]
    reground -.-> meta
    style meta fill:transparent,stroke:#58a6ff,stroke-width:2px
    style outer fill:transparent,stroke:#3fb950,stroke-width:2px
    style inner fill:transparent,stroke:#d29922,stroke-width:2px
    style reground fill:transparent,stroke:#bc8cff,stroke-width:2px,stroke-dasharray:4 4

One piece of the machinery is deliberately not a loop. A loop is feedback — something exists, gets checked, gets corrected. Before any loop runs, a GROUND Ladder governs the input side: how much the writer reads before drafting, widening from the page's declared dependencies toward the whole corpus only on a named signal that the evidence so far is insufficient. The loops correct what came out; the ladder disciplines what goes in — every authoring cycle enters through it. (Details under Key Features.)

The first three loops mirror the "software factory" playbook for AI-assisted coding; the fourth exists because knowledge, unlike code, keeps decaying after you ship it. The full argument for this design is in the companion article: The Knowledge Factory.

Everything else in this README — the commands, the tools, the feature list — hangs off this map.

What makes this different

There are plenty of takes on Karpathy's LLM Wiki idea now. After reading the popular implementations, three things here are genuinely rare — and they are the bet:

• Authoring guidelines that evolve themselves (the meta loop) — something I haven't found in the other implementations. When the same review failure keeps recurring, the system drafts a fix to its own writing rules. The idea is borrowed from Self-Harness and Microsoft SkillOpt. Where the fix touches prose an LLM judge scores, it has to clear a blind A/B against a regression set first; tool code, fact corrections and structural moves are verified deterministically instead, and I can waive the A/B by directing a change myself, with the reason recorded. So it isn't only the wiki that improves over time, but the rules that build it. (Thirteen weeks of that ledger on a private instance: 107 amendments adopted, 72 rejected, 18 deferred. Of the 107, 16 cite a measurement, 40 were adopted on a recorded waiver, and 51 are changes the A/B doesn't cover or that record no measurement. Over the same window, 44 of the 70 defect classes that ever got a fix came back afterwards, and five of them produced 451 of the 823 recurrences (55%); the rule amended most often has been rewritten 19 times and still fires. The loop is still experimental; I'm measuring whether it earns its keep rather than claiming it's solved.)
• A full newsroom, not just "an agent" (the outer loop) — plenty of tools wrap one agent around your notes, and a few add a verifier. Here authoring and review sit in different hands, the review is held to an editorial rubric drawn from real craf