Guides · Skills

NCI skills

Skills are short SKILL.md files that tell an agent how to think when it touches NCI. They live under skills/ in the repo and ship as a single shared knowledge layer for any client that loads them.

What a skill is

A skill is a plain Markdown file with frontmatter. The frontmatter names the skill and decides whether the host can auto-invoke it; the body is a short protocol the agent follows when the skill is in scope.

skills/
  README.md
  <skill-name>/
    SKILL.md

Install

Install both shipped skills from git with npx skills:

nci

nci-answer-quality sets disable-model-invocation: true so the protocol is not paraphrased when it fires. nci leaves invocation flexible for onboarding judgment.

Hosts that understand the open Agent Skills shape each discover SKILL.md from their own paths — there is no single global folder. The host map below is a reference; check each product’s docs when they ship an update.

Where each host looks

Host	Typical locations	How it behaves
Cursor	`.cursor/skills/…`, `.agents/skills/…`, plus user-level paths listed in Cursor’s skills docs	Complements always-on Rules (`.cursor/rules/`). Skills are loaded when the task matches the description — not on every message.
Claude Code	`~/.claude/skills/` (user), `.claude/skills/` in the repo (project)	Directory name becomes a slash command (for example `/deploy` for a folder named `deploy`); see Claude Code skills.
Claude Desktop / Claude apps	Managed in the app (Skills library, connectors, synced packs) — not the same on-disk layout as Claude Code	For repo-local workflows, many teams still keep a `skills/` tree in git and point agents at it via prompts or filesystem access.
OpenCode	`.opencode/skills/`, `~/.config/opencode/skills/`, and compatibility paths such as `.claude/skills/` (see OpenCode skills)	Skills are usually invoked through OpenCode’s `skill` tool so only metadata loads until the agent needs the body.
Antigravity	Workspace `.agent/skills/` and user `~/.gemini/antigravity/skills/` (see Google’s Antigravity skills codelab)	Same bundle layout as other hosts: one folder per skill with `SKILL.md` at the root.
Codex	Repo and parents `.agents/skills/` and user `~/.agents/skills/` (see OpenAI’s Codex skills docs)	Loads name + description first, then the full `SKILL.md` only when the skill is invoked (progressive disclosure). Per-skill enable/disable lives in `[[skills.config]]` blocks in `config.toml`.

Shipped skills

Skill	When it fires	What it forces
`nci`	Fresh context, first NCI use in a repo, or “how do I use NCI?” before any query.	Verify CLI (`nci --version`), `nci init -y`, `nci index`, then read the compact primer (`nci://primer/agent` or bundled `PRIMER.md`) and run `active-package` → `evidence`.
`nci-answer-quality`	On any “using nci” question, type-level explanation request, or strict-correctness ask about composition operators.	Signature-first reasoning: list decision symbols, fetch their declarations through NCI, derive the conclusion from the signature, show one wrong alternative with its rejection, end with a confidence line.

The skills directory is not part of the CLI surface — nci does not read or write it. It is a sibling to NCI that any agent runtime can mount.