How This Wiki Works¶

Tidemill's knowledge base is an LLM wiki: a persistent, interlinked set of markdown pages that an LLM agent builds and maintains as new sources arrive. It is rendered by MkDocs like the rest of the docs.

Why a wiki and not just documents¶

The usual pattern for LLMs and documents is retrieval: point the model at a folder, let it find relevant chunks at query time, generate an answer, throw the answer away. Nothing accumulates. Every question re-derives the same understanding from scratch.

This wiki inverts that. Knowledge is compiled once and then kept current. When a new source arrives — a competitor's pricing page, a billing engine's changelog, a market report — the agent reads it and integrates it: entity pages get updated, contradictions with older claims get flagged, cross-references get rewritten, the catalog and log get appended. A single source typically touches several pages.

The result is a compounding artifact. The cross-references already exist, the contradictions have already been surfaced, and the synthesis already reflects everything read so far.

The tedious part of running a knowledge base was never the reading — it was the bookkeeping. That is the part the agent absorbs.

The three layers¶

┌──────────────────────────────────────────────────────────────┐
│  SCHEMA        .claude/skills/llm-wiki/SKILL.md              │
│                Structure, conventions, and the three ops.    │
│                Read by the agent, co-evolved by humans.      │
└───────────────────────────┬──────────────────────────────────┘
                            │ governs
┌───────────────────────────┴──────────────────────────────────┐
│  WIKI          docs/wiki/                                    │
│                entities/  concepts/  index.md  log.md        │
│                Written and maintained entirely by the agent. │
│                Published to the docs site.                   │
└───────────────────────────┬──────────────────────────────────┘
                            │ compiled from
┌───────────────────────────┴──────────────────────────────────┐
│  RAW SOURCES   docs/wiki/sources/                            │
│                Immutable. The agent reads, never edits.      │
│                In git, but excluded from the published site. │
└──────────────────────────────────────────────────────────────┘

Raw sources are the input layer: captured articles, pricing pages, reports, release notes. They are immutable — the agent reads from them and never rewrites them. They live in git so any clone can re-read them, but MkDocs exclude_docs keeps them off the published site so Tidemill does not republish third-party material.

The wiki is the agent's output. Humans read it; the agent writes it. Two page types carry most of the weight:

Directory	Holds	Example
`entities/`	A company, product, or system	ChartMogul
`concepts/`	A domain idea that recurs across sources	Net Revenue Retention

Longer-form analysis lives in the existing Research section, which acts as the wiki's synthesis layer — narrative pages that argue a position by linking out to entity and concept pages rather than restating their facts. Architecture and Development docs are in scope for the same maintenance and lint passes, but they remain hand-authored specifications: they describe what Tidemill does, not what the wiki has learned.

The schema is .claude/skills/llm-wiki/SKILL.md. It defines page structure, naming, frontmatter, link conventions, and the workflow for each operation. It is what makes the agent a disciplined maintainer instead of a generic chatbot, and it is referenced from AGENTS.md so every session picks it up automatically.

The three operations¶

Ingest — a new source is added to sources/ and the agent processes it: reads it, discusses the takeaways, writes or updates the affected entity and concept pages, records any contradiction with what the wiki already claimed, updates the catalog, and appends to the log.

Query — a question is answered from the wiki, starting at the catalog rather than by re-reading raw sources. When an answer turns out to be worth keeping — a comparison, a newly-noticed connection — it gets filed back as a page so the exploration compounds instead of evaporating into chat history.

Lint — a periodic health check: contradictions between pages, claims a newer source has superseded, orphan pages nothing links to, concepts referenced everywhere but lacking a page, and drift between the catalog, the MkDocs nav, and what is actually on disk.

Navigating¶

Catalog is content-oriented — every page with a one-line summary, grouped by category. It is the entry point for both humans and the agent; a query starts here and drills down.

Log is chronological — an append-only record of every ingest, query, and lint pass. Entries use a fixed prefix so the history stays greppable:

grep "^## \[" docs/wiki/log.md | tail -5

Conventions¶

Frontmatter on every page: title, description, type, tags, updated. description is the one-liner reused verbatim in the catalog.
Links are relative and end in .md — they resolve on the docs site and on GitHub.
Claims carry citations. A factual claim that came from a source links to that source's page or URL. Unsourced claims are marked as inference.
Dates are absolute (2026-07-23), never "last month".
Contradictions are recorded, not silently resolved. When a new source disagrees with an existing claim, both are kept with dates until a human decides.

The full, enforceable version of these rules lives in the schema.