FAQ

What is "Skillerr"?

"Skillerr" names the reference CLI package (npm i -g skillerr), not the project. The project is the Open .skill Protocol — an open specification and portable .skill format so AI apps and agents can author, inspect, version, and run skills interoperably. Full breakdown of every name involved (protocol, format, npm scope, CLI, wire identifiers): Naming. Site: skillerr.com.

What is the Open .skill Protocol?

An open specification for portable AI skills: a sealed .skill package with typed inputs, workflow, pinned knowledge, redacted journey provenance, integrity digests, and optional mint attestation. Conforming hosts adopt the same create / ingest / trust / run lifecycle.

What is the skillerr package?

skillerr is the reference implementation CLI (bin: skill) — one conforming surface for create, inspect, and run. The protocol is not limited to this package; hosts may integrate the @skillerr/* protocol libraries or ship an independent conforming port.

How do I use it?

Install once, then talk to your AI — paste a prompt that points the agent at create, inspect, load, or handoff. You review and approve releases.

npm i -g skillerr

See Getting started.

Why would an app implement the protocol?

So skills are interoperable: one package moves between tools with inspectable digests, TrustView before execute, continuity handoffs, and release completeness gates — instead of proprietary chat dumps or free-form markdown alone.

How is this different from SKILL.md?

Structured package + digests + mint + continuity handoff + release compile gates. skill to-skill-md (single file, quick) is a lossy adapter; skill export-skill (a full folder) preserves frontmatter (license/compatibility/metadata/allowed-tools) and scripts//references//assets/.

How is skillerr different from npx skills add or a skills directory?

They solve different problems and work together. Installers like vercel-labs/skills and directories like skills.sh get a skill onto your machine; they don't check its integrity or provenance before you run it. skillerr is the verification layer: skill ingest wraps a standard Agent Skills folder in a typed contract, an integrity seal, and provenance, so you can inspect and verify a skill, whether you wrote it or installed it, before running it.

Continuity vs release?

• Continuity — work handoff between agents or hosts (partial OK, not mintable).
• Release — complete reusable skill or compile_refused.

Who creates skills?

Agents create (declared host required). Humans install once, prompt the agent, review and approve release compilation. Declared host/model provenance is self-reported unless a verified issuer is configured.

Can I set SKILL_HOST=human?

No. Denylisted hosts (human, cli, shell, manual, …) cannot mint. Do not fake a human host.

How do I create a skill?

Ask your agent (see Getting started). The agent sets SKILL_HOST and runs init → propose → checkpoint or compile --approve --mint. Prefer exact human-approved section bodies. See Create a skill and Agents.

How do I ingest or run a skill?

Ask your agent to inspect first, then validate, then dry-run. See Ingest a skill.

Where do I publish?

Share the .skill file directly. Optional local transparency log: skill registry …. Hosted registries are out of scope for the protocol.

How do I get a public, verifiable provenance URL?

Run skill publish <file.skill>. It seals a release and anchors the digest to the public Sigstore Rekor transparency log, then prints an independently-checkable search.sigstore.dev URL. Zero setup: the public log needs a signing key but no login/account/OIDC, so a per-user key is auto-generated on first run and reused after. Only the digest and skill id are logged, never your content. Rekor entries are permanent and world-readable, so never publish a secret skill. This is a public provenance record, not a marketplace. See Trust and security.

Do I need a blockchain?

No. skillerr does not mint tokens, issue NFTs, or move value, and none of that is required to author, verify, or run a skill today. A ledger anchor is a documented, unimplemented roadmap item, one optional PermanenceAnchor kind among several, never required. See Trust and security for the full extensibility story.

Is this ready to use?

Yes, for real use today — digests, validation, inspect-before-run, and the runtime capability gate are all real, tested code. The protocol spec is versioned 1.0.0 (Stable) — future changes go through the open RFC process, not silent revisions.

For production trust, configure a real Ed25519 issuer key (skill keygen + --signer-key) so mints earn verified_issuer trust — the bundled zero-setup key is for trying the CLI, not for shipping. A configured key alone isn't enough, though: verified_issuer also requires real agent-runtime evidence (a session id or runtime markers), or the mint fails loudly with a clear error rather than silently downgrading. See the OSS repo's Key ceremony for the full generate → mint → pin walkthrough, and What is verifiable for exactly what a signature proves either way.

Do local / offline agents work?

Yes. Ask your agent to set SKILL_HOST / provider / model / deployment for Ollama, LM Studio, llama.cpp, and similar. Runtime dry-run stays local; model invocation in execute mode needs a host adapter.

What does unsigned mean?

Untrusted. Execute refuses unless the host explicitly allows untrusted run.

Can a seal prove which local model wrote the skill?

No. A seal proves which key signed which claims. Local LLMs can still misreport authorship. See Trust and security.

Claude Desktop claims .skill on macOS — is that Skillerr?

No. On macOS, Claude Desktop may register the .skill extension for its own Agent Skills format (a zip with SKILL.md inside). Skillerr .skill files are a different sealed package (typed contract, digests, mint, assets) — same extension, different format.

What to do: Use the CLI (skill inspect ./file.skill) or your AI agent to identify the file. In Finder, use Open With → your editor or terminal instead of double-clicking if the wrong app opens. Do not assume every .skill file is Claude-native or Skillerr-native without inspecting.