---
name: research-workflows
description: "Umbrella for research discovery, monitoring, scraping, knowledge-base building, market/event data research, and ML paper writing."
version: 1.0.0
author: Hermes Agent
license: MIT
metadata:
  hermes:
    tags: [research, arxiv, literature-review, scraping, monitoring, knowledge-base, papers, polymarket]
---

# Research Workflows

Use when the task is to find, monitor, extract, organize, or write research-grade information. This umbrella keeps discovery tools and downstream writing/knowledge workflows in one place so agents do not miss the pipeline shape.

## Triage

- **Literature discovery**: arXiv search by keyword/author/category/ID; collect links, titles, abstracts, and relevance notes.
- **Blog/feed monitoring**: recurring or one-shot monitoring of blogs and RSS/Atom feeds; summarize changes and avoid duplicate noise.
- **Web scraping / difficult extraction**: handle bot detection, dynamic content, structured extraction, and reproducibility.
- **Knowledge-base construction**: build/query an interlinked markdown KB from authoritative sources.
- **Research paper writing**: move from idea/design/experiments to NeurIPS/ICML/ICLR-style drafts with templates and checklists.
- **Insurance / regulated-market comparison**: compare existing policy documents with current public offers, competitor AVB/factsheets, legal levers, and negotiation arguments. For Swiss private-insurance meetings and cyber-policy comparisons, use `references/swiss-private-insurance-comparison.md`; for Helvetia AVB update/Update-Garantie/Kassensturz household-insurance benchmark work, also use `references/swiss-helvetia-avb-update-kassensturz.md`. For Swiss life-insurance/Säule 3a exit questions (Rückkauf, Prämienfreistellung, Umwandlung, representative pushback), use `references/swiss-life-insurance-policy-exit.md`.
- **Congressional trading / public official portfolios**: distinguish official disclosure facts from modeled tracker estimates; use `references/congressional-trading-disclosures.md` for STOCK Act caveats, House/Senate sources, tracker links, and useful Quiver endpoints.
- **Swiss product/shopping comparison**: when comparing physical products for fit, do not stop at shop availability and headline dimensions. Extract the product identifiers (MPN/GTIN), current Swiss shop offers, direct product links, material/compatibility, outer dimensions, *usable/effective dimensions* (e.g. cooking surface), and user constraints. Treat user-corrected constraints such as material (e.g. Chromstahl/Edelstahl only) or effective usable area as first-class ranking criteria.
- **Swiss online product search / price comparison**: when finding a specific product across Galaxus, digitec, BRACK, Nettoshop, and Swiss specialist shops, extract MPN/GTIN first, use Toppreise as an aggregator when direct shop access is blocked, and format a concise price/availability table; see `references/swiss-online-product-search.md`.

## Workflow

1. Define the research question and acceptance criteria: breadth scan, focused evidence, or publication-quality artifact.
2. Gather sources with the correct subsection; prefer primary sources and retain URLs/IDs.
3. Extract into a structured intermediate form: citation metadata, claims, evidence, dates, caveats, and follow-up queries.
   - For shopping/product comparisons, separate **headline dimensions** from **functional dimensions** (e.g. outer size vs usable cooking surface), and rerank when the user clarifies what actually matters. Material constraints (Chromstahl/Edelstahl vs coated aluminum/cast iron) must be filtered early, not handled as an afterthought.
4. Synthesize with explicit uncertainty and avoid unsupported extrapolation.
5. For recurring monitoring, schedule via cron/blogwatcher only after the one-shot query is proven useful.
6. Preserve scripts, templates, and reproduction commands in the final artifact.

## Preserved source packages

Former narrow skills were copied under `references/absorbed-packages/<skill>/` before archiving: `arxiv`, `blogwatcher`, `llm-wiki`, `research-paper-writing`, `web-scraping`, and `polymarket`.