Roadmap: restructure into a public, agent-agnostic skills repo (owid/skills)

[lucasrodes]

This repo packages OWID skills for the public, but it drifted: Claude-centric framing, four overlapping plugins, staff-only skills mixed in, and new development happening in the private staff repo instead. Meanwhile the ecosystem consolidated around the open Agent Skills standard (SKILL.md), now natively supported by Claude Code, OpenAI Codex, Gemini CLI, Cursor, GitHub Copilot and ~40 other agents — with orgs like Anthropic and Sentry publishing a single skills repo that doubles as a Claude Code plugin marketplace.

Decision: restructure + rename in place, don't archive & recreate. GitHub renames preserve history, stars, PRs, and auto-redirect old clone/raw URLs; a fresh repo gains nothing since the skill content is good — only the packaging was dated.

This issue tracks the phased plan. Details per phase are collapsed below.

Phases

• Phase 1 — Restructure the repo → Restructure as a public, agent-agnostic skills repo (single 'owid' plugin) #7
• Phase 2 — Rename to owid/skills & re-point the staff repo
• Phase 3 — Discoverability & public docs
• Phase 4 (optional) — Hosted read-only MCP

Phase 1 — Restructure the repo (PR #7)

• One public plugin, owid, bundling search-charts, fetch-chart-data, joining-data, fact-check-article (added in Add fact-check-article skill to owid-data-web #6), owid-catalog — top-level skills/ layout (cross-agent convention, same pattern as anthropics/skills).
• Drop owid-general (uv/duckdb: internal working practices) and owid-general-staff (datasette, create-chart: staff-only) — both live exclusively in the private staff repo now.
• Versionless plugins: every commit to main is a release (official marketplace guidance; removes the forgot-to-bump footgun).
• Public-first README with per-agent install: Claude Code marketplace, npx skills add (75+ agents), manual .agents/skills copy.
• AGENTS.md as the real contributor guide, CLAUDE.md symlinked to it; add missing Apache-2.0 LICENSE.
• Remove manage-agent-skills.py TUI (superseded by npx skills); trim prerequisites script to uv, jq, duckdb.
• Follow-up: rebase Add fact-check-article skill to owid-data-web #6 (fact-check-article) onto the new layout. Done — Add fact-check-article skill to owid-data-web #6 merged in parallel and Restructure as a public, agent-agnostic skills repo (single 'owid' plugin) #7 has been rebased on top of it; fact-check-article now lives at skills/fact-check-article/ and is part of the owid plugin.

Phase 2 — Rename to owid/skills & re-point the staff repo

Rename (needs team sign-off on the name):

• GitHub rename owid-claude-plugins → owid/skills. Old URLs (clone, web, raw) auto-redirect, so nothing breaks; then update self-references in README and install-prerequisites-macos.sh cosmetically.
• Update repo description + topics (agent-skills, claude-code, mcp, ai-agents, open-data, …) for search.

Fix the flow direction with the private staff repo (currently a diverged superset of this one — that's the real structural problem):

• Staff repo stops vendoring the four public plugins; staff install both marketplaces (public owid/skills + private staff).
• Truly internal plugins (owid-general-staff, owid-general, chart-animator, vibe-apps, customize-charts) remain private — decided.
• Public-relevant skills developed there in the future should land here first.

Phase 3 — Discoverability & public docs

Someone googling "owid mcp" or "owid claude" should land somewhere official. Currently they find nothing.

• Page on docs.owid.io (or ourworldindata.org) — "Using Our World in Data with AI assistants" — explaining what's available and linking here. This is the canonical landing page for those searches.
• Submit/verify listing on skill directories: skills.sh, officialskills.sh, VoltAgent/awesome-agent-skills; check gh skill discoverability.
• Announce internally so staff stop installing from the private repo for public skills.

Phase 4 (optional) — Hosted read-only MCP

Skills + CLI tools are the right fit for read-only public data (cheap in context, portable, no server to run) — that consensus holds. But MCP still matters for two things skills can't cover:

1. Agent surfaces without a shell (claude.ai web, ChatGPT, mobile) can't curl — a hosted MCP is the only way in.
2. The "does X have an MCP?" discovery channel — comparable data providers already ship one: World Bank Data360 MCP, Google Data Commons MCP (hosted on Google Cloud).

If pursued: a thin hosted server exposing exactly what the web skills do (search + fetch chart data/metadata), no auth, read-only — the Data360 MCP is the model. Separate project & repo; explicitly not a blocker for phases 1–3.

Open decisions

• Final repo name — proposal: owid/skills (matches anthropics/skills, getsentry/skills, openai/skills)
• Who owns the docs.owid.io / website page (Phase 3)
• Whether Phase 4 (hosted MCP) happens at all, and where it would run