Skills Catalog

The skills/ directory holds public, generic agent skills. Each skill is a directory with a SKILL.md (frontmatter: name, description) following the agentskills.io format.

Install them into your agents (Claude Code, Codex, Cursor, …) with:

npx skills add mappedinfo/metis

Visual skills also need the shared profiles installed into the project they work in. Skill installation does not copy the repository's sibling profiles/ directory. See the profile installation instructions for the metis-os profiles install command and source-checkout entry.

Admission criteria

A skill belongs in metis only if all of the following hold:

1. Generic. It encodes reusable craft, not one person's identity, preferences, private projects or private data. Personal overlays belong in USER.md, never in a skill.
2. Self-contained. It works without access to any private repository, vault or service. External dependencies must be public and named.
3. Honest. Instructions must not ask the agent to invent evidence, fabricate citations, or misrepresent capability.
4. Scoped. One capability per skill. If a description needs "and also…" twice, it is two skills.
5. Writable description. The frontmatter description says when to use the skill and when not to, in the third person, without marketing language.
6. Artifact-anchored. The required artifacts frontmatter names concrete, verifiable deliverables — a downloaded PDF, a figure, a deck, a runnable package or a planning note with acceptance criteria. The description or Output contract explains the output and how to check it.

Every skill also belongs to exactly one of seven capability categories — acquire · analyze · compose · visualize · present · build · meta. Categories are capability domains, not scenarios: a paper or a deck is a combination of skills. Shared style (palette, typography) lives once in profiles/ and is referenced with uses:. Each output keeps the profile for its audience and format; a journal figure and a presentation deck can share palette and font roles while retaining different type sizes and export defaults. Profile IDs resolve in the current project root's profiles/ directory.

Skills also carry two controlled tag axes in frontmatter — domain (1–2 values, e.g. writing, figure, data-analysis) and action (1–3 values, e.g. create, review, transform). The controlled vocabulary lives in skill-contract.json. The authoritative admission criteria live in skills/README.md.

Validation

Contributors run npm run skills:check from the repository root. The repository adapter builds on the published metis-atlas@0.1.0 scanner and writes .atlas/skills.json, .atlas/graph.json and .atlas/audit.json. It checks controlled vocabulary, nonempty artifacts declarations, and uses: references to skills and profiles. It does not judge output quality or establish that a task was completed.

npm run skills:check -- --domain paper exports a domain view after auditing the full collection. Native npx metis-atlas scan in version 0.1.0 does not provide these metis contract and profile dependency checks.

Catalog

Skill	Category	Domain	Description
scansci-pdf	acquire	literature	Academic paper acquisition via the scansci-pdf MCP server (17 tools): search, verify, cite, download (Apache-2.0, vendored with provenance)
metis-router	meta	workflow-meta	Composes skills for multi-stage tasks: one primary + ≤2 supports per phase, profiles for shared style, artifacts per phase

Every skill is a bounded SOP: a ## When to stop section names the budget and termination criteria, ## Requirements lists [B]/[A]-labeled verifiable items, and unbounded-loop phrasing fails skills:check. See CONTRIBUTING.md for the template.

The router bundles a reusable requirements ledger. At close-out, each accepted task and skill requirement has done, blocked or n/a plus evidence or a reason. n/a applies only outside the accepted scope; it cannot waive a required user outcome. Planning-only work checks the plan and identifies downstream execution as unperformed. Budget exhaustion reports completed items, unfinished items and the smallest unblock condition; it does not establish task completion. The declaration lint does not verify execution ledgers or their evidence.

Contributing

Open a PR with one skill per PR. Expect review on admission criteria 1–6 above. See skills/README.md for the tag vocabulary and format details.