Research Skills – Claude Code Skills for Academic Work

What it is – A set of five Claude Code “skills” (plug‑in style command‑line tools) that help researchers automate parts of the academic workflow, especially for AI‑related papers. The skills cover:

Skill	Main purpose
medical‑imaging‑review	Build a rigor‑checked, citation‑integrity‑driven literature review for medical‑imaging AI.
paper‑slide‑deck	Turn any document into a set of AI‑generated, stylised slide images (17 visual themes).
research‑proposal	Generate a PhD‑level research proposal in a Nature‑Reviews style, with verified references.
scholar‑slides	Create an editable, vector‑based slide deck where every equation, table, figure and citation is faithfully reproduced from the source.
lit‑search	Perform exhaustive, time‑windowed literature retrieval with a measurable recall report.

---

How it works

1. Installation – Each skill is a folder that you copy into ~/.claude/skills/. Some skills (scholar‑slides, lit‑search) ship an install.sh that creates a Python virtual‑env, installs Node/npm packages and any required system fonts.
2. Invocation – Skills are triggered by a slash command (e.g., /lit-search) or a CLI call such as paper-slide-deck content.md --style watercolor. The README lists the exact trigger strings for each skill.
3. Core philosophy – All writing‑related skills share a citation‑integrity discipline: stable cite‑keys, a claim‑ledger, and an audit script that never silently passes a manuscript. The integrity guardrails are documented in docs/citation-integrity-core.md.

---

Quick look at each skill

1. medical‑imaging‑review

• Target – Systematic‑style reviews of AI for medical imaging.
• Workflow – Choose a route (narrative, method‑survey, scoping, systematic). The tool scaffolds a project (init_review_project.py) with a timestamped config, claim ledger, bibliography and checklist files. An audit script (audit_manuscript.py) cross‑checks every citation against the ledger and reports failures, warnings or “not assessed”.
• Outputs – Markdown manuscript, references.bib, claim_ledger.csv, and a PRISMA‑style checklist.

2. paper‑slide‑deck

• Target – Visual, share‑ready slide images (not editable). Useful for quick overviews or social media.
• Key steps – Detect figures in a PDF, map each figure to a full‑page slide, generate an image via the Gemini‑3‑pro‑image API, then merge into PPTX or PDF.
• Limitation – Math and data are rasterised; not suitable for a defense where precise numbers must stay editable.

3. research‑proposal

• Target – Draft PhD proposals or research plans (English or Chinese).
• Process – Five‑phase pipeline: requirements → literature collection (with verification) → outline approval → section‑by‑section writing → final Markdown output.
• Guardrails – Every reference is checked against DOI/PMID/arXiv or a Zotero library; unverifiable entries are flagged.
• Typical output – 2‑4 k word proposal, 25‑50 verified references, 3‑5 figure suggestions, ready for pandoc conversion.

4. scholar‑slides

• Target – High‑fidelity, editable slide decks for journal clubs, conference talks, etc.
• Features – Vector equations via KaTeX, real tables, exact figure crops, bilingual support, and an integrity QA gate that flags missing or unverified content.
• Outputs – Reveal.js HTML deck, vector PDF, and an editable PPTX with speaker notes.
• Known issue – Figure localisation works best on single‑column layouts; two‑column papers may need manual confirmation.

5. lit‑search

• Target – Exhaustive literature retrieval with a transparent recall metric.
• Approach – Multi‑source crawling (OpenAlex, PubMed, Crossref, etc.) with full pagination, citation‑closure, and a “coverage_report.md” that lists what was not retrieved.
• Depth tiers – Different cost/precision levels; you can preview time and cost with uv run lit depths.
• Outputs – dois.md, references.md (formatted in several citation styles), references.bib, PRISMA summary, coverage report, and a CSV queue for human screening.

---

Who might use this?

• Researchers preparing systematic reviews or proposals who want reproducible, auditable workflows.
• Graduate students needing quick, citation‑safe slide decks for seminars.
• Lab managers who want a deterministic way to collect all papers on a topic and quantify search completeness.
• Anyone using Claude Code who prefers a plug‑in model rather than writing custom scripts.

---

License

MIT – free to use, modify, and redistribute.

---

Bottom line – The repository is a genuine collection of open‑source tools that automate scholarly writing and presentation tasks while enforcing strict citation integrity. It is squarely in the AI‑assisted research‑automation space.