A self-maintaining knowledge graph on agentic AI, compiled from technical literature, blog posts, and industry/consulting papers. Scanners pull sources in; they're compiled into an interlinked wiki that compounds over time. It follows the Karpathy "LLM-wiki" pattern and is an Obsidian vault, so wikilinks and graph view work natively.
It serves a senior AI/tech exec persona — signal over textbook: what's genuinely
new, why it matters, and the "so what." Voice and node-quality rules live in
wiki/_templates/STYLE.md; architecture and operating
conventions live in CLAUDE.md.
| Path | What it holds |
|---|---|
raw/ |
Immutable source captures — raw/arxiv/, raw/blogs/, raw/news/, plus per-year censuses. Scanners only ever ADD here; never edited. Not part of the graph. |
wiki/ |
The maintained graph: concepts/, entities/, sources/, comparisons/, overview.md, and _templates/ (node structures + style guide). |
index.md |
Catalog of every wiki page — read first, then drill in. |
log.md |
Append-only timeline of every ingest / lint pass. |
CLAUDE.md |
The schema: architecture, ingest/query/lint operations, graph-density rules. |
scanners.md |
Source list + agents-centric focus filter the scanners read. |
skills/ |
The automations — arxiv-scanner/, blog-scanner/, kg-curator/ (see skills/README.md). |
briefings/ · trackers/ · radar-notes/ |
On-demand exec deliverables, one doc type per folder (created on first use). |
- Census — every matching paper as metadata in
raw/arxiv/<year>-census/. Exhaustive, not in the graph. - Browsable — abstract-only auto-notes in
wiki/sources/<year>-census/, keyword-linked, filtered out of the global graph view. - Deep — full-text reads with analysis (
depth: full-text), first-class graph nodes.
Three scanners feed the graph, each a different layer of truth: arXiv = what's
true (research), news = what happened (events), blogs = what it means /
who thinks what (attributed opinion + debates). After a scan, the ingest
operation writes/refreshes source notes, integrates findings into the relevant
concept/entity pages (no duplicates; contradictions flagged inline), promotes
node-worthy references to their own linked nodes, then updates index.md and
appends to log.md. The kg-curator skill periodically tends the graph — mining
new nodes, enriching citation-ranked sections, normalizing to templates, and linting
for orphans, broken links, and stale framing.
Two other operations: query (answer from the graph with citations to the pages used; file durable answers back as new pages) and lint (health-check: contradictions, stale claims, orphans, missing cross-references).
Kebab-case filename slugs; Obsidian [[slug|Display]] wikilinks (the links are the
graph). Every node has ≥1 inbound and ≥1 outbound link — no orphans — and every
relationship is reciprocal (back-link on both ends). Concept and entity pages are
synthesized, compounding articles grounded in full text, not abstracts; source notes
are one per ingested item. Full node structure and style are in wiki/_templates/.
57 concepts · 157 entities · ~1,700 source notes · 2025 + 2026 agents-centric
censuses (~23k paper records). See index.md for the live catalog and log.md for
the running history.