wiki-ingest is the general ingestion skill of the obsidian-wiki set. It reads a source, extracts the knowledge in it and integrates that knowledge across the wiki, creating or updating concept, entity, reference and project pages with cross-links.
It handles append, full and raw-staging modes, large folders and git repositories, scanned images, academic papers, web URLs and meeting or voice-note transcripts. Source text is treated as data, so instructions inside a document are never followed.
What You Get
- Distilled concept, entity and reference pages with cross-links and frontmatter
- Provenance markers on inferred and ambiguous claims
- A manifest and cache that skip sources already ingested
- A batch plan for large folders and a guard check before parallel runs
- Project reference pages for web URLs, with duplicate checks and stubs for pages that fail to load
- Meeting pages with decisions, commitments and open questions
Common Use Cases
- Turning a folder of notes and PDFs into a connected wiki
- Saving a web article into the right project’s references with one link
- Pulling meeting transcripts from a connected tool into dated wiki pages
What Makes It Unique
It aims to integrate knowledge across existing pages, and warns against writing a summary per file. It has a stated trust boundary for source content, a cache keyed on content so unreliable timestamps do not force rework, event-time dates taken from the source, and a size ceiling per page so Obsidian can still open the vault. Paper ingestion keeps figures, equations and results tables that text extraction drops.
Under The Hood
The skill is instructions plus four reference files. It reads source files and the vault, writes markdown pages and edits the manifest, and runs the obsidian-wiki command line tool for cache checks, batch planning, an ingest guard, code structure extraction, timestamps and a locked memory sync. For web URLs it fetches pages with defuddle, a web fetch tool or firecrawl, and falls back to the Jina Reader service by calling it with curl, which sends the URL to that service. Meeting pulls call a connected meeting tool. Long PDFs can run an external PageIndex script that calls a language model. Optional steps use the qmd tool and git. Source content is treated as untrusted data.