A Claude-maintained knowledge wiki built on Qvantum TechDoc output (QG & QH heat pumps, Modbus). This is the one-stop handover: what it is, what's inside, what you're inheriting, and exactly how to keep feeding it new documentation.
The raw documentation is the single source of truth. Claude derives a curated, human-friendly wiki from it. The schema file tells Claude how. You (the curator) drop sources and ask questions — the agent does all the writing, linking, and bookkeeping.
raw/Immutable TechDoc bundles (DITA-OT HTML + SVG figures), one folder per product line.
wiki/Sources, entities, concepts & queries — cross-linked Obsidian pages with provenance back to the GUIDs they came from.
CLAUDE.mdThe agent's playbook: page formats, naming rules, ingest & lint workflows, hard rules.
Products covered: Qvantum QG (ground-source, QG-7 & QG-14, incl. the Sweden edition with SmartControl & BBR), Qvantum QH-175 (hydronic unit), and the Modbus integrator register map. Query pages include the Swedish customer FAQ, the feature-availability matrix, the 111-term terminology harvest, and the 15-phase installer checklist.
Every page under wiki/ — cross-linked, sourced, and graph-clean. Open the vault in Obsidian and start at the index.
wiki/index.md · wiki/log.md · wiki/products.mdThe whole approach is portable. Drop the setup spec into a new repo, point Claude at it, and it reproduces the same structure and workflows for a different document stream.
WIKI-AGENT-INSTRUCTIONS.md · CLAUDE.mdThe provenance manifest (.ingest.yaml) gives hash-based delta detection, so re-published bundles only touch the wiki pages that actually changed.
build_ingest.py (WIKI-AGENT-INSTRUCTIONS §6)A per-market status dashboard built from one templated PowerPoint per country. Markets fill the template, drop it in input/, run one script.
dashboard/build.py · /update-mdu-dashboardraw/, archive/ and the chat-history hook are gitignored or live outside the repo — the checklist captures exactly what a plain git clone leaves behind.
TRANSFER.mdRestore-point .zip snapshots in archive/wiki snapshots/ roll the wiki back before any big change. Chat transcripts live in archive/chat history/.
rm -rf wiki && unzip "<snapshot>.zip"raw/New product → a new folder raw/<PRODUCT>/. Updated product → drop the new bundle over the existing folder (same GUIDs, higher revision).
Open the vault as the working directory and say "ingest the new bundle in raw/". Claude reads the source, then discusses the key takeaways with you before writing anything.
It zips a restore point, regenerates .ingest.yaml, writes the source page, and creates/updates the entity & concept pages — every claim cited, every page carrying provenance back to its source GUIDs. Target: 5–15 pages touched, no stubs.
For a re-published bundle, Claude diffs file hashes against the manifest and only re-reads and rewrites the pages the changed topics feed into. Cost scales with what changed, not the bundle size.
Every ingest, delta, and lint pass is recorded in wiki/log.md (newest on top) and reflected in wiki/index.md — so the catalog is always current.
| File | What it is |
|---|---|
wiki/index.md | Master catalog — the entry point for every page in the wiki. |
CLAUDE.md | The agent's day-to-day playbook: page formats, naming, ingest & lint workflows, hard rules. |
WIKI-AGENT-INSTRUCTIONS.md | Transferable setup spec — reproduce this whole pattern on a new project / document stream. |
TRANSFER.md | Checklist for moving the vault to another machine without losing gitignored pieces. |
README.md | Human-facing intro: layout, roles, workflows at a glance. |
dashboard/ | The MDU per-market status dashboard (PowerPoint → build.py → index.html). |