Skillget

claude-academic-workflow

mcp
Security Audit
Fail
Health Warn
  • License — License: NOASSERTION
  • Description — Repository has a description
  • Active repo — Last push 0 days ago
  • Low visibility — Only 5 GitHub stars
Code Fail
  • eval() — Dynamic code execution via eval() in docs/assets/mathjax/MathJax.js
  • exec() — Shell command execution in docs/assets/mathjax/MathJax.js
Permissions Pass
  • Permissions — No dangerous permissions requested

No AI report is available for this listing yet.

SUMMARY

Public Claude Code workflow for academic research: 32 skills, 4 sub-agents, 3 hooks, MCP catalog (stdio + claude.ai OAuth), Notion + Telegram task orchestration. Fork-and-adapt; cross-platform (Windows + macOS). Personalization via ~/.claude/state/personal_config.json.

README.md

claude-academic-workflow

An academic-research workflow for Claude Code: skills for reading papers, running literature reviews, auditing bibliographies, drafting preregistrations, writing referee reports and R&R responses, assembling replication packages, compiling LaTeX, iterating TikZ figures, and a Quarto reveal.js slide system with render-time quality gates. It was built for quantitative marketing and economics; most of it transfers to any empirical field.

Everything in this repository, including both example decks and their figures, is AI-generated with Claude Code, as a proof of concept for what an agent-built research workflow looks like. Generated content is the responsibility of whoever uses it: verify citations, numbers, and claims before relying on them, the same way you would verify a research assistant's first draft.

Two live example decks, rendered by these skills and published on GitHub Pages:

  • Research talk: a seminar deck on a light theme, with staged reveals, jump buttons into the appendix, a paginated reference list, and a segmented progress bar.
  • Teaching lecture: a classroom deck on a dark theme, with teaching blocks, a worked example, discussion prompts, and agenda tracking.

Both are static pages that make zero network requests at display time; the index links them with keyboard shortcuts. Their sources are in examples/, and they render out of the box against the plain starter theme in slide-tooling/ (the live copies use a private theme built on the same hooks).

Quickstart

git clone https://github.com/ericluo04/claude-academic-workflow
cd claude-academic-workflow
cp -R skills/* ~/.claude/skills/
cp agents/tikz-reviewer.md ~/.claude/agents/
mkdir -p ~/.claude/assets && cp -R slide-tooling ~/.claude/assets/quarto-yale

Then read SETUP.md: it names every path and helper the skills assume (the Crossref contact address, the PDF helper, the Overleaf glob) and how to adjust each one. CLAUDE.md is the author's working global configuration, shared as an example. The three style/house.md files in the slide skills are stubs for your own taste: author line, closing-slide wording, palette rationale, density calibration.

Skills

Skill What it does
reading-papers Looks up and reads a specific paper from a link, DOI, title, or vague description, across marketing, economics, psychology, and CS venues.
litreview Finds, ranks, and synthesizes the literature on a topic across Semantic Scholar, OpenAlex, and arXiv, then reads the top hits with parallel subagents.
bibcheck Audits a .bib file entry by entry against canonical metadata and writes a corrected copy; catches wrong years, mis-cited authors, and hallucinated entries.
preregister Drafts a registry-ready preregistration (AsPredicted, OSF, AEA RCT) with clarity flags and placeholders instead of invented content.
council Spawns five independent critic subagents in parallel on any target, then a synthesis pass that ranks findings by how load-bearing they are.
review-paper Runs a full referee review with six agents in parallel, triaged CRITICAL/MAJOR/MINOR against a named target journal.
referee-response Drafts an R&R response letter where every claimed change is located and read in the manuscript before a location pin is written.
replication-package Assembles a journal-ready replication archive and scans it for secrets, PII, and absolute local paths first.
compile-latex Compiles with latexmk, auto-detects the engine and bib backend, and emits a ranked error report with file:line attribution.
tikz-iterate Compiles a TikZ figure, rasterizes it, has the tikz-reviewer agent actually look at the image, and iterates until approved.
research-talk Authors a Quarto reveal.js deck for a seminar, conference talk, or job talk: assertion titles, staged reveals, a deep appendix for questions.
teaching-lecture Authors classroom lecture decks built for engagement: worked examples, discussion prompts, checks for understanding, projector-sized figures.
slide-review Renders a deck, screenshots every slide in a real browser, and reviews the pictures for overflow, contrast, broken figures, and a weak argument.
course-site Builds the semester course website the lecture decks hang off, and publishes it to GitHub Pages.

agents/tikz-reviewer.md is the adversarial visual critic tikz-iterate loops on. slide-tooling/ holds the machinery the slide skills share: the staging filter, the fit and staging gates, the offline checker, a starter theme, and vendored MathJax and Inter; its README documents all of it.

Things you may not know

A few of the higher-leverage discoveries from building this, none of them obvious from the docs.

  • Claude Code's agents mode runs several named sessions side by side, each with its own context, which is how the parallel-subagent habit in CLAUDE.md gets practical: one session drives, others research or build in the background. The part that is easy to miss: clicking the left arrow on an agent re-activates that existing session with its context intact, so you can hand a running agent more work instead of starting over.

  • When a university mail account cannot integrate with anything, draft in a personal Gmail via the Gmail MCP and copy the draft out of Gmail's web editor into the university client. Gmail-to-client copy-paste preserves formatting, where terminal-to-client copy-paste mangles spacing and line breaks (observed on Windows; other platforms may differ). The agent writes and revises the draft; you paste and send.

  • macOS Terminal can tell you when a long run finishes: in the profile settings, the audible bell plus the option to have new output announced by voice means you can walk away from a render or a scrape and get called back. Cheap, and it changes how you use long-running agents.

  • MCP integrations worth knowing:

    • Zotero: the agent searches your library, checks retractions before you cite, pulls metadata, and reads the attached PDFs, so "is this in my library and is it still good" becomes one question.
    • Playwright: a real browser the agent drives, for scraping, for form-filling, and for visually reviewing rendered pages and slides rather than trusting the markup.
    • Gmail, Google Calendar, Google Drive: drafting mail (see above), checking availability before proposing meeting times, and reading shared documents without downloading them by hand.
    • Notion: reading and writing pages and databases, useful when a lab or a course already lives there.
    • Hugging Face: searching models and datasets and reading repo files, for the ML side of a research pipeline.
    • A scholarly-search connector: semantic search over the academic literature as a tool call, which the reading and lit-review skills use as one rung of their source ladder.

License and credit

MIT, see LICENSE. The skills adapt ideas from several public workflows (Pedro Sant'Anna, Scott Cunningham, Chris Blattman, Claes Bäckman, and others); ATTRIBUTION.md traces each one and covers the vendored Inter (OFL) and MathJax (Apache 2.0) copies.

Reviews (0)

No results found