Research & Analysis Workflows

Use AFTER literature-review and BEFORE ingest-source to obtain the actual PDFs for the A+B graded sources. Auto-downloads Open-Access PDFs into the shared library ( /pdf/ .pdf — the bibkey IS the filename) and writes a manual-download worklist (input/bibliography/acquisition-todo.md) for everything paywalled or bot-blocked, so the user can fetch originals via university VPN. Re-run to reconcile manually-downloaded files. Triggers on "acquire the sources", "download the literature", "get the PDFs", "Quellen beschaffen", "lade die Literatur herunter". NOT a search skill (that is literature-review), NOT an ingest skill (that is ingest-source).

Add ONE document (a PDF) directly to the shared master library — outside the project ingest flow. Auto-discovers the bibliographic metadata, VERIFIES it against Crossref/OpenAlex (never guessing), derives keywords cheaply from the PDF's docinfo + first-page abstract (no full read), computes the canonical bibkey, places the PDF at /pdf/ .pdf and appends the entry to /references.bib. Stops and asks when metadata cannot be verified. Triggers on "add this PDF to the library", "add to the master bibliography", "put this straight in the library", "füge das Dokument der Bibliothek hinzu", "nimm das PDF in die Bibliothek auf". NOT ingest-source (which reads a source into a project's wiki), NOT acquire-sources (which downloads a review's worklist).

Use to draft a book chapter, article section, or grant exposé from synthesized wiki content. Settles the argument architecture with the user first, then drafts section by section — each section's argument agreed before its prose is written. Pulls from stable synthesis pages and source pages, writes to output//.qmd with proper citations. Never drafts from memory — always from the wiki.

Use to check the shared library, the current wiki, and all registered projects for accumulated drift — un-indexed PDFs, scans without a text layer, PDF bloat, project-bib keywords not yet merged into the master, cross-project bibkey collisions, lint regressions from out-of-band edits. Wraps hooks/driftcheck.py, the same state-triggered check that runs automatically at session start. Triggers on "drift report", "check everything", "was steht an", "prüf mal alles", "ist alles konsistent", "check the library". NOT wiki-lint (single-project structural lint — this composes it), NOT semantic-wiki-review (LLM content audit — this is deterministic only).

Use AFTER writing-research-plan has produced a ready plan (status=ready for hermeneutic projects, status=pre-registered for quantitative/mixed). Reads input/ideas/ -plan.md, creates TodoWrite items, runs a pre-ingest acquisition gate (acquire-sources) so originals are on disk before any source-ingester runs, and works tasks via subagent dispatch (source-acquirer, source-ingester, analyst, drafter) with two-stage review. Does not execute anything outside the plan.

Use when a research project reaches completion — manuscript finalized, peer-review cleared, ready for publication / submission / archive. Runs the final checklist across rendering, BibTeX, wiki, data, reproducibility, archive, and follow-up grant planning.

Use for researching German and European funding programs (DFG, ERC, VolkswagenStiftung, Gerda Henkel, Fritz Thyssen, bilateral DE-IL) for projects in theology, biblical archaeology, digital humanities. Produces a DFG-style research-status draft, funding ranking, panel recommendation, and BibTeX.

Use when a manuscript (book chapter, article, grant) is drafted and ready for structured review. Runs a two-stage peer-review workflow — constructive then adversarial — with discipline-specific checklists (CONSORT/STROBE + theology/archaeology/DH-specific).

Use to create a new research project from the template via conversation — no terminal, no cp commands. Asks for project name, parent directory, methodology, discipline, and languages; copies the template tree using the Write tool; patches CLAUDE.md frontmatter; optionally initialises git. Designed to work in Cowork (no shell access) as well as Claude Code.

Use when the user asks for a content review of the wiki — contradictions between pages, outdated claims overtaken by newer sources, weakly supported assertions, missing cross-references that the structural linter cannot detect. Distinct from wiki-lint (structural, deterministic, CI-gated).

Use when starting any research conversation — establishes how to find and use research skills. Requires Skill tool invocation before ANY response, including clarifying questions. Governs the full research workflow from idea to publication.

Use to build and analyse the knowledge wiki as a graph. Runs scripts/wiki-to-graph.py (from the research project template) to export an interactive self-contained graph.html (open in any browser, no install) plus graph.json / graph.graphml — and answers questions against the live wiki via deterministic query sub-commands (neighbors, path, god-nodes, bridges, relations, search, stats), grounded in the result — most-connected pages (god nodes), entities that bridge otherwise-unconnected sources, relation types and confidence (inferred vs grounded), clusters and weak spots. Triggers include "build/show me a knowledge graph", "map the wiki", "graph view of my research", "which entity connects or bridges my sources", "what are the most connected or central pages", "find surprising connections", "where are the gaps or weak links in the wiki", "export the wiki to Gephi/yEd". For frontmatter and wikilink validation use wiki-lint; for content and claim audit use semantic-wiki-review.

Use AFTER brainstorming-research has produced an approved design doc. Converts the design into a pre-registered research plan with falsifiable hypothesis, concrete task list, and verification criteria. No execution happens without this plan.