Imported from chanmainvest/ingoldwetrust_library (llm-wiki/AGENTS.md). Install upstream with npx skills add chanmainvest/ingoldwetrust_library --skill llm-wiki. Copyright stays with the author.

AGENTS.md — LLM-Wiki (In Gold We Trust)

This is an LLM-Wiki for the In Gold We Trust report corpus, following Andrej Karpathy's LLM-Wiki pattern.

Obsidian is the IDE; the LLM is the programmer; the wiki is the codebase.

Purpose

Turn the raw, year-by-year report Markdown in ../markdown/ into a persistent, compounding knowledge base of cross-cutting concept pages. The reports are organized chronologically; the wiki is organized thematically. A question about "central bank gold buying" should resolve to one concept page that cites every relevant chapter across all 20 years, not 20 separate file reads.

The three layers

1. Raw sources — ../markdown/<year>/. Read-only. The agent never edits these.
2. The wiki — this folder (llm-wiki/). Written and maintained entirely by the agent. Concepts live in concepts/; navigation lives in index.md and log.md.
3. The schema — this file (AGENTS.md). Defines structure and operating rules.

File conventions

• Concept pages live in concepts/<topic>.md, e.g. concepts/central_bank_gold.md. Filenames are lowercase with underscores.
• Each concept page opens with a one-line definition, then a Sources section of citations: - [Year — Chapter](../../markdown/YYYY/NN_title.md). (Concept pages live in concepts/, so reaching markdown/ at the repo root requires two ../.)
• Every concept page MUST cite at least two source chapters before it is created. Single-source "concepts" stay as notes in log.md until a second source appears.
• Pages cross-link each other with relative paths: [inflation](central_bank_gold.md).

Navigation files

• index.md — the catalog. The agent reads this FIRST to locate relevant pages. Two sections: the source corpus (what exists) and the concept index (what has been compiled).
• log.md — append-only chronological record. Every ingest/query/lint action adds one entry with the header format: ## [YYYY-MM-DD] ingest|query|lint | <short title>

The agent loop

Ingest (adding knowledge from a source)

1. Read the source chapter in ../markdown/YYYY/NN_title.md.
2. Identify which existing concept pages it informs; update each with a new citation and any new nuance. Touch as many pages as warranted — don't be lazy.
3. If the source introduces a concept not yet in the wiki AND a second source already touches it, create concepts/<topic>.md.
4. Otherwise, log the nascent concept in log.md under a "candidates" note.
5. Append an entry to log.md.

Query (answering from the wiki)

1. Read index.md to find relevant concept pages.
2. Read those pages; synthesize an answer grounded in their citations.
3. If the answer required real synthesis across sources, offer to file it back as a new concept page or update an existing one. Knowledge should compound.

Lint (health check)

1. Scan all concepts/*.md for:
   • Broken citation links (source file moved/deleted).
   • Orphan pages (no incoming links from other concepts).
   • Pages with only one citation (promote or fold back into log.md candidates).
   • Contradictions between pages.
2. Append a summary of findings to log.md.

Operating rules

• Never fabricate citations. Every [Year — Chapter] link must point to a real file in ../markdown/. If unsure, verify the path exists.
• Read before write. Always read a concept page before editing it.
• One concept per page. If a page grows past ~300 lines, split it.
• Dates in log.md use ISO format (YYYY-MM-DD) so the history is unix-greppable.