I’ve been using Claude Code (Anthropic’s terminal coding agent) in my climate and weather work, and over time I built up reusable “skills” and reviewer “agents” tuned for AI-assisted scientific research. I’ve cleaned them up, documented them, and released them as the “ai-science-toolkit” under the MIT License.
A potentially useful part of the the toolkit are the domain-expert reviewer agents. These are review-only personas you point at a result or a draft section, and they encode the questions a skeptical colleague might ask, e.g.:
- attribution-reviewer: is the counterfactual defined, the baseline stated, are the claims over-reaching?
- scicomm-reviewer: is the message/framing appropriate for public-facing materials?
You can run one, or several in parallel as a panel on any document or piece of code you are drafting.
A few skills that may be of interest:
- commit-batch: helps manage git repos… batches the current working tree into logical, single-concern commits, then commit and push if asked
- figure-review: reviews a plot for publication-readiness, e.g. flags jet/rainbow, checks for colorblind-safe palettes, assess whether uncertainty is missing, considers axis/caption gaps
- lit-review: searches/synthesizes literature across Zotero/arXiv/bioRxiv/Scholar tools
- overbaked: audits a document, plan, or code for over-engineering, verbosity, and scope creep
- write-new-skill: creates new Claude Code skills with proper structure
To Install (as a Claude Code plugin, one line from within Claude Code CLI):
/plugin marketplace add dgilford/ai-science-toolkit /plugin install ai-science-toolkit@ai-science-toolkit
Or you can clone and install pieces individually (just grab what you need). The README explains how to do both!
This set of tools grew out of my own workflows in detection & attribution, sea level change, and tropical cyclone analysis (mainly thinking about research → operations pipelines)… the tools most likely still reflect that. It’s still relatively early on in production (just release v1.0.0) and the tools remain strongly shaped by my own workflow and habits. I’m sharing here in case they’re useful to others doing geoscience in Python, and because I’d value feedback from people doing similar work.
Do you find this useful, and do you have any feedback on these tools? Thanks in advance.