A grimoire — a book of spells — for the RLadies+ community and global team: a collection of Agent Skills that help everyone follow the RLadies+ guide, from writing in the RLadies+ voice, through blogging, translation, and social media, to coordinating the work the global team does behind the scenes. The name is a nod to the rest of the witch-themed RLadies+ tooling — glamour, cloak, and spellbind.

Each skill (a spell) packages the relevant parts of the guide into focused, on-demand guidance that an LLM agent (e.g. Claude) loads only when it's actually needed. The set is modelled on the rOpenSci Skills repo.

RLadies+ is a community non-profit, so the work clusters around communications, content, governance, and the website rather than code. The strongest skills are the ones where an agent can actually produce, standardise, or review an artifact — a post, a blog draft, a translation, a review assignment. Tasks that are purely clicking through an external portal (rotating an API key, triaging mail, operating a phone line, nonprofit filings) stay in the guide as runbooks; they aren't agent skills.

rladies-voice and rladies-brand are the two foundation skills every content skill builds on: rladies-voice covers the voice and tone (the do/don't rules and the drafting pass that keeps copy from reading AI-generated), rladies-brand covers naming, the non-negotiable inclusive-language rule, visual identity, and accessibility. The rest compose on top of both.

All thirteen skills from the project's evaluation are now available. rladies-voice was added afterward as the RLadies+ voice skill — the identity-level do/don't rules that used to live in rladies-brand's Voice section now live there, alongside the drafting/AI-tell material.

The rladies-* skills above are context skills — an agent loads them when a task is relevant and they supply the rules. On top of them sits a set of action skills that a team member invokes explicitly to run a review on an artifact as a pre-human gate: point one at a draft, a branch, or a PR and it loads the matching context skill, executes the checklist (and any real commands — blog-lint expectations, alt-text scans, inclusive-language greps, live CI output on a PR), and emits a consistent Blockers / Warnings / Nits report with a location and a suggested fix for each finding. They never approve, merge, schedule, or publish — they hand a human a punch list.

They all emit the same tiered report (defined in run-brand-check), so reviewers get familiar output whichever one they run.

The repo is packaged as a single Claude Code plugin (grimoire) that bundles the skills, and doubles as a one-plugin marketplace:

grimoire/
├── .claude-plugin/
│   ├── marketplace.json   # the marketplace catalog
│   └── plugin.json        # the plugin manifest
└── skills/
    └── <skill-name>/
        ├── SKILL.md        # lightweight guidance, loaded when the skill triggers
        └── references/     # deeper detail (templates, schemas, walkthroughs), on demand

The fourteen context skills: rladies-voice (voice, foundation) · rladies-brand (identity, foundation) · rladies-blog-post · rladies-social-posts · rladies-translate-page · rladies-abstract-review · rladies-branded-assets · rladies-chapter-event · rladies-rocur · rladies-website · rladies-partnership-eval · rladies-chapter-setup · rladies-funding · rladies-community-moderation.

The action (run-*) skills: run-brand-check (shared core) · run-blog-review · run-social-review · run-translation-review.

Each skill follows the Agent Skills format: SKILL.md carries the lightweight guidance loaded once the skill triggers, while references/ holds the deeper detail the agent reads only when a task calls for it. The run-* skills are pure SKILL.md (no references/) — they orchestrate the context skills rather than adding new material.

In Claude Code, add this repo as a marketplace and install the plugin:

/plugin marketplace add rladies/grimoire
/plugin install grimoire@grimoire

/plugin marketplace update pulls newer versions later, and /plugin opens the manager to enable, disable, or uninstall.

Most of these skills distil guide.rladies.org; when the guide changes, update the matching skill. Part of rladies-voice is guide-distilled (the identity-level do/don't rules) and part is a practical companion authored alongside it (the AI-tell material) — its SKILL.md says so up front. New skills should follow the same pattern: a tight SKILL.md with a trigger-rich description, deeper material in references/, and the RLadies+ voice rules from rladies-voice and inclusive-language rules from rladies-brand throughout.

Each content skill carries a sources.yml manifest pinning the exact guide.rladies.org pages it condenses, modelled on the same tooling in rOpenSci Skills. Two monthly GitHub Actions keep the two from drifting apart silently:

• drift-detect (tools/drift-detect.R) checks whether any pinned page changed since a skill's ref, and files/updates a drift-labelled issue when one has.
• coverage-audit (tools/coverage-audit.R) checks the inverse: whether a guide page exists that no skill's manifest references yet (and isn't in tools/coverage.yml's out_of_scope list), and files/updates a coverage-labelled issue.

Adding or updating a skill's guide references should also mean adding or updating its skills/<skill>/sources.yml. tools/reconcile-prompt.md is the guardrail an LLM follows when drafting a sync PR off a drift finding — it defaults to no change and only edits when the upstream diff changed guidance the skill is responsible for conveying.