Turn today's science breakthroughs into Instagram-ready carousels — grounded in real sources, zero hallucinated citations.
A Claude Code skill that fetches live research from arXiv, PubMed, ScienceDirect, Nature, ScienceDaily & more, then generates publication-ready carousel slides, captions, hashtags, and full academic citations.
AI-generated science content has a 40-91% citation hallucination rate. Fake DOIs, invented journal names, non-existent authors. ScienceToolkit solves this by fetching real sources first, then generating content grounded entirely in verified material.
/science # auto-discover today's trending topic
/science quantum computing # or pick your own
One command triggers a 7-step pipeline:
Discover Topic ─→ Load Rules ─→ Fetch 7 Sources ─→ Generate Content ─→ Validate ─→ Write File
│ │ │ │
RSS + WebSearch arXiv, PubMed, 5-7 slides, 5 format
14-day diversity ScienceDirect, caption, hashtags, checks
tracking ScienceDaily, APA citations (PASS/FAIL)
Phys.org, Nature,
Ars Technica
Each run produces a dated Markdown file in output/:
output/
├── 2026-03-16-crispr-gene-editing.md
├── 2026-03-16-solar-fuel-conversion.md
└── topic-log.json # 14-day diversity tracking
Every file contains:
| Section | Details |
|---|---|
| Slides | 5-7 carousel slides with hooks and cliff-hangers |
| Caption | 400-600 words, keyword-optimized, < 2100 characters |
| Hashtags | Exactly 5, strategically chosen |
| Sources | Full APA citations with DOIs, URLs, peer-review labels |
| Images | Source image URLs with license status |
Slide 1: Bacteria Just Lost Their Best Defense
What if superbugs could be tricked into disarming themselves?
Gardner, J. P., et al. (2023). The James Webb Space Telescope Mission. Publications of the Astronomical Society of the Pacific. DOI: https://doi.org/10.1086/723378 [Published in: PASP, 2023]
- 7-channel parallel fetching — arXiv, PubMed, ScienceDirect, ScienceDaily, Phys.org, Nature News, Ars Technica
- Auto-topic discovery — scans RSS feeds for trending topics, picks the most cross-referenced
- 14-day diversity tracking — won't repeat the same topic within two weeks
- Format validation — 5 mechanical checks before output (caption length, hashtag count, slide count, label format, citation completeness)
- Write-and-warn — always saves the file, prepends warnings if validation fails
- Peer-review labeling — every source tagged as
[Published],[Preprint], or[News article] - Zero config — no API keys, no external dependencies, runs entirely on Claude Code's native tools
- Have Claude Code installed
- Clone this repo:
git clone https://github.com/d7rocket/ScienceToolkit.git cd ScienceToolkit - Run:
claude # opens Claude Code in this directory - Generate a carousel:
/science
That's it. No API keys, no npm install, no Docker.
ScienceToolkit/
├── .claude/skills/science/
│ ├── SKILL.md # 7-step pipeline (the engine)
│ └── prompts/system.md # content grounding rules
├── examples/
│ └── output-sample.md # output format contract
├── output/ # generated carousels live here
└── .planning/ # architecture & research docs
| Component | Role |
|---|---|
| Claude Code Skill System | Orchestration — /science slash command |
| WebSearch / WebFetch | Live source fetching with dynamic filtering |
| arXiv REST API | Academic papers (via export.arxiv.org) |
| PubMed E-utilities | Peer-reviewed medical research |
| ScienceDirect (via WebSearch) | Peer-reviewed Elsevier journal articles |
| RSS (ScienceDaily, Phys.org, Nature) | Trending topic discovery |
- Source-first architecture — fetch before generating, never the reverse
- Single focal finding — each carousel tells one story, not a topic survey
- Character ceiling, not word count — Instagram truncates at ~2100 chars, so we count characters
- Inline glosses — every technical term gets a plain-English explanation in parentheses
- Cool professor tone — calm, confident, accessible, no jargon gatekeeping
MIT
Built with Claude Code