convez.ai
MCP ServerProdutividadeAvançado
ericluo04

Claude Academic Workflow

ericluo04/claude-academic-workflow

Fluxo de trabalho acadêmico completo com 32 skills e orquestração Notion+Telegram

Ver no GitHub
InicialPluginScore · de 0 a 100

Um sistema MCP completo para pesquisa acadêmica em Claude Code, com 32 skills especializadas para manuscritos, revisão de literatura, gerenciamento de referências e orquestração de tarefas via Notion e Telegram. Focado em pesquisadores quantitativos que usam LaTeX, Beamer e ferramentas acadêmicas padrão, com suporte cross-platform (Windows e macOS).

Instalação

$npx convez-ai ericluo04/claude-academic-workflow

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
26
Forks
3
Acompanhando
0
Saúde
42%
Último commit
há 2 dias
Versão
—
Issues abertas
0
Contribuidores
2

Casos de uso

  • Redação e estruturação automática de artigos científicos em LaTeX
  • Busca integrada de literatura (arXiv, Semantic Scholar, OpenAlex) com deduplicação
  • Gestão de referências bibliográficas e auditoria de citações
  • Orquestração diária de tarefas de pesquisa via Notion e Telegram

Veredito

Solução robusta e opinada para pesquisadores quantitativos — bem pensada, documentada e battle-tested em uso diário. Requer configuração inicial e familiaridade com LaTeX/Beamer, mas elimina fricção significativa no ciclo de pesquisa acadêmica.

README

A tradução para o português deste README ainda está na fila.

claude-academic-workflow

A complete academic-research workflow for Claude Code — 32 skills, 4 sub-agents, 3 hooks, 6 stdio MCPs plus the claude.ai connector catalog, Notion + Telegram task orchestration, cross-platform (Windows + macOS).

What this is

This repo packages a working Claude Code configuration for academic researchers — particularly those in quantitative marketing, economics, or other quantitative social-science fields. The skills cover the full research lifecycle: drafting and reviewing manuscripts, lit-review across arXiv / Semantic Scholar / OpenAlex, citation auditing, slide-making (Beamer + PowerPoint), TikZ figure iteration, replication-package assembly, cross-language code audits, and day-to-day task orchestration through Notion + Telegram.

The setup is battle-tested in daily PhD-research use. It is not a demo. That said, it is opinionated: it assumes you write papers in LaTeX, draft slides in Beamer, keep a .bib per project, and track tasks in Notion. The skills ship clean of personal data — every project name, page ID, and home path resolves through ~/.claude/state/personal_config.json at runtime — so a friend can fork the repo, drop their own IDs into the config file, and pick up the workflow without inheriting the original author's projects.

Quick start

Five commands to a working setup. Side-by-side per OS.

Windows (PowerShell):

git clone https://github.com/ericluo04/claude-academic-workflow.git
cd claude-academic-workflow
.\scripts\install.ps1
notepad $env:USERPROFILE\.claude\state\personal_config.json
# In a separate step, set TELEGRAM_BOT_TOKEN in your shell or GitHub Actions secrets

macOS (bash/zsh):

git clone https://github.com/ericluo04/claude-academic-workflow.git
cd claude-academic-workflow
./scripts/install.sh
$EDITOR ~/.claude/state/personal_config.json
# In a separate step, set TELEGRAM_BOT_TOKEN in your shell or GitHub Actions secrets

Then in Claude Code: /daily-brief --hours 4 to verify the loop end-to-end. Full walk-through in SETUP.md.

What's inside

At a glance:

ComponentCountLocationIndex
Skills32skills/SKILL_INDEX.md (4-way lookup), skills/README.md (one-liners)
Sub-agents4agents/agents/README.md
Hooks3hooks/hooks/README.md
Personal config (template)1skills/_config/skills/_config/README.md
Settings templates2config/config/README.md
Docs11docs/see Documentation below
Helper scripts6scripts/see Scripts below

The skill listing below is grouped by task. The full alphabetical index with trigger phrases and composition pointers is in SKILL_INDEX.md.

Drafting and writing (skills/)

  • /draft — section drafting in your configured voice (intro, methods, limitations, etc.)
  • /cite — resolve a DOI / arXiv / title to a .bib entry via Zotero
  • /litreview — multi-source lit search (arXiv + Semantic Scholar + OpenAlex), deduped and ranked
  • /preregister — draft an AsPredicted / OSF / AEA preregistration

Reviewing and auditing (skills/)

Slides and figures (skills/)

Task orchestration (skills/)

  • /daily-brief — score and rank open tasks, push top N to Telegram
  • /capture — process inbound Telegram replies into Notion task updates
  • /log-todo — mid-session capture into the Notion Tasks DB
  • /notion-log — append a dated entry to a Notion project page
  • /notion-meeting-notes — extract action items from a Notion meeting page into the Tasks DB
  • /task-pulse — read-only ad-hoc questions against the Tasks DB

Replication and pipeline hygiene (skills/)

Meta (skills/)

One-page 4-way lookup table (by task, by trigger phrase, by category, by composition): SKILL_INDEX.md. Full per-skill index with one-line descriptions: skills/README.md.

Sub-agents (agents/) — specialist reviewers spawned by the slide-quality skills, restricted to read-only tools.

Hooks (hooks/) — harness-lifecycle scripts, all fail-open so a bug never blocks Claude Code.

Personalization — every skill resolves personal data at runtime from ~/.claude/state/personal_config.json (Notion IDs, Telegram chat ID, Overleaf root, project list, voice-reference .tex). Template + field-by-field documentation in skills/_config/. To adapt the workflow to your own projects: docs/adapting.md.

Requirements

  • Claude Code (latest)
  • git, GitHub CLI (gh)
  • Python >= 3.12 (+ uv for the stdio MCPs)
  • Node.js >= 20 (for the OpenAlex / Playwright / GitHub MCPs)
  • LaTeX: MiKTeX on Windows, MacTeX (or BasicTeX) on macOS

Per-OS install commands and recommended optional tooling: docs/clis.md.

MCPs

Three integration shapes — install commands, OAuth flows, and per-skill consumption table in docs/mcps.md.

1. Stdio MCPs (6, open-source, run locally via claude mcp add)

  • arxiv — paper search / download / citation graph (no secrets)
  • semantic-scholar — paper / author / citation-graph queries (free API; optional key for higher quota)
  • openalex — bibliographic graph, author disambiguation (free "polite pool" email)
  • zotero — read-only library lookup for /cite (API key)
  • playwright — browser automation for OAuth flows and verification (no secrets)
  • github — PR / issue / repo ops (local binary, GitHub PAT)

2. HTTP / OAuth gateway MCPs (1, CLI-added)

  • notion — workspace I/O; backbone of /log-todo, /notion-log, /capture, /task-pulse, /notion-meeting-notes, /daily-brief

3. claude.ai web-portal connectors (OAuth, browser-based — no claude mcp add needed)

Installed by clicking "Connect" in the claude.ai Settings → Connectors catalog. Tools then appear inside Claude Code with the mcp__claude_ai_<Provider>__<action> prefix. The catalog evolves — verify the live set on the connector page.

  • Google Drive — Drive file search; used ad hoc by /cite and reference lookups
  • Gmail — email read/triage and draft creation; planned /email-triage skill
  • Hugging Face — search models / datasets / papers / spaces; used by /litreview on ML topics
  • Scholar Gateway — institutional-access bridge for paywalled content; best-effort assist to /cite, /bibcheck, /litreview
  • Microsoft 365 — Outlook / Calendar / Teams / OneDrive; commonly blocked by university tenants — substitute via Gmail forwarding + iCal, see docs/outlook-gmail.md
  • Lifestyle / travel — Resy, StubHub, Booking.com, Expedia, Tripadvisor, Trivago. Optional; useful for conference-trip logistics from /log-todo-style asks.

Full per-connector breakdown (scopes, why-useful, status): docs/mcps.md.

Workflow architecture

+-------------------------+        +-----------------------------+
|  Claude Code (skills)   | <----> |  Personal config (gitignored)|
|  ~/.claude/skills/      |        |  ~/.claude/state/            |
|  ~/.claude/agents/      |        |    personal_config.json     |
|  ~/.claude/hooks/       |        +-----------------------------+
+-------------------------+
            |
            | reads / writes
            v
+-------------------------+
|  MCPs                   |     6 stdio: arxiv, semantic-scholar,
|  6 stdio + 1 HTTP       | --> openalex, zotero, playwright, github
|  + claude.ai connectors |     1 HTTP gateway: notion
+-------------------------+     claude.ai catalog: Google Drive, Gmail,
                                Hugging Face, Scholar Gateway, M365,
                                Resy/StubHub/Booking/Expedia/Tripadvisor/
                                Trivago
            |
            | task data + diary entries
            v
+-------------------------+        +-----------------------------+
|  Notion workspace       | <----> |  lan-daily-brief (separate  |
|  (Tasks DB, project     |        |  repo) — Telegram bot +     |
|  pages, weekly agenda)  |        |  GitHub Actions cron        |
+-------------------------+        +-----------------------------+
                                              |
                                              v
                                       Telegram bot
                                       (morning brief,
                                       capture replies)

Skills live under ~/.claude/ and read all personal data from ~/.claude/state/personal_config.json. The orchestration repo runs scheduled GitHub Actions that call the same Notion API and push to a Telegram bot — see orchestration/README.md for the high-level shape and the link to the companion repo ericluo04/lan-daily-brief.

Documentation

Eleven docs under docs/, one per concern. Start with platforms.md if you are setting up a fresh machine.

DocCovers
platforms.mdSide-by-side Windows / macOS / Linux paths, package managers, and toolchain matrix
clis.mdRecommended CLIs with install commands and version checks
mcps.mdFull MCP catalog: stdio installs, HTTP gateway, claude.ai connectors, token refresh
tex-setup.mdMiKTeX / MacTeX / TeX Live for slide and replication skills
notion-setup.mdNotion Tasks DB + Weekly Agenda + project pages layout and integration
telegram-setup.mdBot creation, token handling, chat-ID retrieval
overleaf-dropbox.mdOverleaf ↔ Dropbox sync for the project tree the skills read from
outlook-gmail.mdSubstitute path when Microsoft 365 OAuth is blocked at your university
adapting.mdHow a friend forks the repo and rewires it to their own projects
future-work.mdRoadmap of unshipped skills and patches
attribution-table.mdMaster attribution table (per-skill, per-source, per-license)

Scripts

Helper scripts in scripts/:

  • install.ps1 / install.sh — copy skills / agents / hooks into ~/.claude/, render settings.json from the template, scaffold state/.
  • verify.ps1 / verify.sh — post-install sanity check (MCPs reachable, personal_config fields filled, hook scripts executable).
  • redact-check.py — local pre-commit gate that scans for secrets and personal IDs against scripts/.blocklist.json. CI runs gitleaks over the same surface — see Security.
  • _render_settings.py — template substitution helper used by the installers.

Configuration

Two settings templates in config/, both rendered by the installer into ~/.claude/:

Both target files are user-local and gitignored once installed. Details and the ${HOME} substitution rule: config/README.md.

Security

Three layers, all on by default:

  1. personal_config.json is gitignored and lives only under ~/.claude/state/. Skills surface a "config missing" error and refuse to proceed rather than guess if a field is unset — guessing produces silently wrong Notion writes. Bot tokens and API keys never go in this file; they live in ~/.claude/state/telegram.json (gitignored) or GitHub Actions Secrets. See skills/_config/README.md for the safe-to-share table.
  2. scripts/redact-check.py runs locally before every commit, scanning files against an exact-string blocklist (scripts/.blocklist.json, gitignored; example at .blocklist.example.json) and a pattern set.
  3. CI secret scan — .github/workflows/secret-scan.yml runs gitleaks against .gitleaks.toml on every PR to catch anything the local check missed.

Setup

Full step-by-step in SETUP.md. The installer scripts (scripts/install.ps1 / scripts/install.sh) handle the file copies; the rest is configuring Notion, Telegram, and MCPs in order. To adapt to your own projects: docs/adapting.md.

Attribution

This repo stands on the shoulders of other Claude Code workflow authors who shared their setups publicly. High-level credits below; full per-item attribution with the lines that were borrowed is in ATTRIBUTION.md.

  • Anthropic (anthropics/skills) — /skill-creator (entire skill, including analyzer / comparator / grader sub-agents, eval viewer, and benchmark scripts); the foundation every other skill in this repo was built on
  • Scott Cunningham (scunning1975/MixtapeTools) — /referee2, /blindspot, /bibcheck, /tikz-iterate (concept), /slide-excellence pedagogy lenses, /create-lecture --triage
  • Pedro H.C. Sant'Anna (pedrohcgs/claude-code-my-workflow) — all four sub-agents (slide-auditor, pedagogy-reviewer, proofreader, tikz-reviewer), /seven-pass-review, /preregister, /slide-excellence base orchestrator, /create-lecture base workflow, /audit-reproducibility 5-phase audit, both compaction hooks (pre-compact.py, post-compact-restore.py), and the workflow-as-config-repo pattern itself
  • Chris Blattman (chrisblattman/claudeblattman) — /council, /daily-brief wait-factor + type-balance scoring, atomic-write discipline, /notion-meeting-notes thin-content gate, /litreview --four-axis, /skill-creator catalog-conflict gate
  • Claes Bäckman (claesbackman/AI-research-feedback) — base designs for all five review skills (/review-paper, /review-paper-light, /review-paper-code, /review-pap, /review-grant), including agent counts and role names
  • Zara Zhang (zarazhangrui/frontend-slides) — /academic-slides scaffolding (Phase 0-5 architecture, "Show Don't Tell" preview UX, STYLE_PRESETS convention, viewport-fit invariant, python-pptx ingest)
  • Gabberflast (Gabberflast/academic-pptx-skill) — /academic-pptx (entire skill — SKILL.md, content_guidelines.md, slide_patterns.md, "Structured Argument" mode, ghost-deck test, action titles)
  • Ethan Weber (ethanweber/posterskill) — /posterskill (entire skill — paper + project-website ingestion, single-file React-via-CDN interactive HTML poster architecture)
  • aspi6246 (aspi6246/Claude-Code-Presentation) — /referee-response --five-q, /review-paper and /review-paper-light buried-contribution check
  • Andrej Karpathy (karpathy/autoresearch) — reviewed; broader "Claude Code as iterative research collaborator" framing cited
  • Hugo Sant'Anna (hugosantanna/clo-author) — reviewed; non-overlapping ideas surfaced in docs/future-work.md

License

MIT — see LICENSE. All source repos credited above ship under permissive licenses compatible with MIT redistribution; see ATTRIBUTION.md for per-source license notes.

Status and disclaimer

This is tuned for one researcher's specific PhD workflow (quantitative marketing, Beamer + LaTeX, Notion + Telegram, Overleaf-on-Dropbox). It is a friendly fork-and-adapt: every personalized field reads from ~/.claude/state/personal_config.json, so you can rewire to your own projects without touching skill prose. Bug reports and PRs welcome; large structural reworks are easier as forks.

Ver o original no GitHub

Bom para

Relacionados