Karpathy-style LLM wiki adapted on my Papers Explained blog corpus and more

A compounding research knowledge base optimized for AI agents. This repository serves as a structured, queryable archive of research papers, concepts, and technical notes, adapted from the "Papers Explained" blog corpus and beyond.

The goal of this project is to build a compounding knowledge base that can be queried in plain English. It tracks research on various topics to provide a single source of truth for complex technical domains.

The wiki follows a strictly layered architecture to ensure data integrity and ease of navigation for both humans and agents.

flowchart TD
    INBOX["📁 inbox/\ndrop source docs here"]
    RAW["📁 raw/\ncanonical ingested sources"]
    ART["📁 artifacts/\ngenerated explainers"]

    subgraph SKILLS["⚙️ .agents/skills/"]
        SI["ingest/SKILL.md"]
        SQ["query/SKILL.md"]
        SQI["query-interactive/SKILL.md"]
        SL["lint/SKILL.md"]
    end

    subgraph WIKI["📚 wiki/"]
        IDX["index.md\nmaster catalog"]
        LOG["log.md\nappend-only log"]
        SUM["summaries/"]
        CON["concepts/"]
        ENT["entities/"]
    end

    AGENTS["AGENTS.md\nschema + rules"]

    subgraph HARNESS["🖥️ harness"]
        AG["Antigravity"] 
        CC["Claude Code"]
        CX["Codex"]
        CUR["Cursor"]
        
    end

    INBOX -->|"$ingest"| SI
    SI -->|moves source| RAW
    SI --> SUM
    SI --> CON
    SI --> ENT
    SI --> IDX
    SI --> LOG

    AGENTS -->|"loads rules"| HARNESS

    SQ -->|"reads"| IDX
    SQ -->|"drills into"| SUM
    SQ -->|"drills into"| CON

    SQI -->|"reads"| IDX
    SQI -->|"drills into"| SUM
    SQI -->|"drills into"| CON
    SQI -->|"writes"| ART

    SL -->|"audits"| WIKI

    HARNESS -->|"runs"| SKILLS

• inbox/: Drop zone for documents waiting to be ingested.
• raw/: Canonical home for ingested source documents. Once a file lands here, it should not be modified.
• wiki/: Structured content pages with cross-references and tags.
• artifacts/: Generated interactive explainers (HTML/CSS/JS) from wiki-backed queries; not ingest input and safe to regenerate.
• index.md: Master catalog with one-line summaries for rapid orientation.
• log.md: Append-only ingestion history.
• .agents/skills/: Repo-local workflow skills for ingest, query, query-interactive, and lint.

This wiki is designed to be managed by AI agents (like Antigravity, Claude Code, or Cursor) using the instructions in AGENTS.md.

To add new research to the wiki:

1. Drop the source file (PDF, text, etc.) into the inbox/ directory.
2. Tell the agent: "ingest inbox/<filename>"
3. The agent moves the file into raw/, then summarizes it, creates the wiki page(s), and updates the index/log.

To retrieve knowledge from the wiki:

• Ask the agent: "query the wiki for [topic]" or simply ask a technical question about the contents. The agent will consult the index.md, find relevant pages, and provide an answer with citations.

For a diagram-heavy or interactive explainer saved in the repo (not just chat prose):

• Ask for an interactive or visual explanation from the wiki, or say to save under artifacts/.
• The agent follows .agents/skills/query-interactive/SKILL.md, reads the same wiki sources as a normal query, and writes artifacts/<YYYY-MM-DD>-<topic-slug>/index.html (plus optional CSS/JS).
• Open index.html locally in a browser; figures resolve via paths like ../wiki/assets/....

To maintain wiki health:

• Tell the agent: "lint the wiki" The agent will scan for broken links, ensure the index is up-to-date, and check for logical contradictions across pages.