Factual Research — AI skill for Claude Code

factual-research

Research with your AI agent where the sources can't be made up. factual-research searches the sources each field actually trusts (PubMed for vitamins, NASA NTRS for composites, arXiv for LLMs, ISAP/SAOS for Polish law…), keeps a PRISMA-style record of what was found and why things were excluded, and then **verifies every citation, quote, number and table cell** in the text your agent writes, before you ever read it.

Same idea as [write-as-me](https://github.com/masluny/write-as-me): the LLM you already use (Claude Code, Codex, the Claude app via MCP) does the reading and writing, and a small, deterministic Python tool does the measuring. You get a 0–100 score with ranked, fixable notes, and the agent loops until the score passes.

No API keys needed for the core sources. Everything is cached locally, and nothing is uploaded anywhere.

question ──init──▶ plan.toml (sub-questions, PICO, native queries, domain packs)
          search ─▶ 17 connectors in parallel → dedupe (DOI/PMID/arXiv/title) → auto-screen
          screen ─▶ agent includes/excludes with reasons (→ PRISMA)
          enrich ─▶ retractions (Retraction Watch via Crossref), preprint→published,
                    study design, venue signals, funding/COI
          fetch  ─▶ open-access full text (PMC XML, publisher PDFs, arXiv, NTRS, SAOS, ISAP)
          matrix ─▶ claims × sources, stance + verbatim quote → consensus + GRADE-style certainty
          tables ─▶ evidence table / summary of findings / custom tables, every cell quote-verified
          draft  ─▶ [@key, p. 4 :: "exact source words"]
          verify ─▶ score + fixes  ⟲ until pass
          render ─▶ final text + tables + references (APA, Vancouver, IEEE, …, .docx, HTML article)
          report ─▶ HTML: PRISMA flow, findings, matrix, forest plot, verified text, sources

Install

pip install -e ".[all]"        # core is stdlib-only; [all] = PyMuPDF, python-docx, openpyxl
factual-research doctor --online