Build an LLM-readable wiki from your documents
Teams keep knowledge in files — reports, specs, meeting transcripts, analyses — scattered across formats that neither search nor models read well. The cheapest durable fix isn't a platform. It's a convention: a git repository of converted Markdown, one document per file, browsable by humans (GitHub renders it), retrievable by models, auditable by diff.
1 · Why the Archive preset is the wiki format
The Archive preset optimizes for faithfulness and provenance rather than paste budget:
---
title: "Q1 operations review"
source: "q1-operations-review.docx"
converted: "2026-07-04"
format: "docx"
---- Full front matter — every page knows what it is and where it came from; tooling (site generators, Dataview, your own scripts) can index it.
- Faithful body — nothing truncated harder than the source forced; this is the copy of record, not the paste copy.
- Figures as real files — the .zip's
figures/folder commits alongside, so embeds render in any Markdown viewer.
2 · Conventions that turn a folder into a wiki
- Stable, descriptive filenames —
2026-03-vendor-consolidation-review.md. Filenames are your URLs, your link text, and your retrieval titles. - One index page — a hand-kept
README.mdgrouping documents by topic. Models navigate indexes brilliantly; so do new teammates. - Convert on arrival — a document enters the wiki by being dropped through the converter, so everything inside obeys the same rules (typed tables, figure placeholders, honest truncation). The fidelity report is your admission check: a file that converts at QC 40% needs a better source before it becomes "knowledge".
- Let git be the history — re-convert updated sources onto the same filename; the diff shows exactly what the new version changed, which is often the most interesting knowledge of all.
3 · What each consumer gets
| Consumer | What the wiki gives it |
|---|---|
| Humans | Rendered pages, working search, one canonical copy |
| Chat assistants | Paste-ready pages; front matter answers "what is this?" |
| Workspaces / RAG | Upload or index the folder; heading-aligned chunks + stable anchors if you re-emit with the RAG preset |
| Agents | A greppable corpus with addresses — the docs-md/ pattern, promoted to team scale |
4 · Starting today, with what you have
Batch-drop one project's documents into the converter with the Archive preset, commit the .zip's contents to a fresh repo, write ten lines of README index. That's the whole bootstrap — the discipline (convert on arrival, stable names, re-convert on update) is what compounds.
Convert the first ten pages of your future wiki in one drop.