naturali.ai skills

Agent Skills for building, running and auditing AI agents on naturali.ai.

Each skill is one atomic module taken from the naturali.ai tutorials: the commands that deliver one result, reusable on their own. A module several tutorials share (create an agent, run a generation, settle an approval) is one skill.

Resources are declared in one formation template, naturali.yaml: naturali-deploy-a-formation owns validate, deploy, plan and update, and every other skill adds its resources to that template. Actions that are not state (generating, uploading bytes, searching, approving, starting a run) are curl calls, with the CLI command and SDK method named beside each.

Install

Claude Code:

/plugin marketplace add naturali-ai/skills
/plugin install naturali@naturali-skills

Any agent that reads Agent Skills: copy the directories under skills/ into its skills folder (for Claude Code, ~/.claude/skills/).

Any client that reads Agent Plugins: this repository is one, with plugin.json, the skills and the naturali MCP server in mcp.json.

Every skill expects a nat_sk_… project API key in NATURALI_TOKEN.

Skills

Start with naturali and naturali-deploy-a-formation; each skill's Before you start names the skills that produce what it needs.

Skill	Does
`naturali`	The API, auth, clients and modules at a glance; start here.
`naturali-create-a-project`	A project to deploy into, its id as `PROJECT`.
`naturali-deploy-a-formation`	Write, validate, deploy, plan, update and tear down `naturali.yaml`.
`naturali-enable-naturali-models`	Every managed model, no vendor account or credential.
`naturali-bring-your-own-model-key`	A provider on your own vendor credential.
`naturali-create-an-agent`	An agent, versioned on every change.
`naturali-run-a-generation`	One answer from an agent, in the background or inline.
`naturali-read-a-run`	What a run was asked, called, returned and cost.
`naturali-return-structured-output`	Replies as a validated JSON object.
`naturali-stop-storing-conversation-content`	An agent whose prompts and replies are never stored.
`naturali-give-an-agent-an-http-tool`	An agent that calls an HTTP endpoint.
`naturali-run-tools-in-your-own-code`	A tool your own code runs, the generation paused until it answers.
`naturali-call-a-tool-directly`	A tool tested on its own, with no agent or model.
`naturali-gate-a-tool-with-guardrails`	Small tool calls run, large ones wait for a person.
`naturali-settle-an-approval`	A held action approved, edited or rejected.
`naturali-make-a-file-searchable`	A file turned into a searchable document.
`naturali-search-knowledge`	The passages a question retrieves, ranked.
`naturali-ground-an-agent-in-documents`	An agent that answers from the project's documents.
`naturali-give-an-agent-long-term-memory`	An agent that remembers facts across conversations.
`naturali-let-an-agent-write-memories`	An agent that saves what it is asked to remember.
`naturali-limit-what-an-agent-may-do`	An agent the platform refuses some actions to.
`naturali-connect-a-discord-channel`	An agent answering Discord direct messages.
`naturali-converse-in-a-session`	A multi-turn conversation, read back turn by turn.
`naturali-replay-a-turn`	A bad answer re-answered on its exact history.
`naturali-roll-out-an-agent-version`	A new version served to a share of traffic, then promoted.
`naturali-gate-a-rollout-on-an-eval`	A rollout that promotes only after its eval passes.
`naturali-build-an-eval-dataset`	Test cases, hand-written and from real traffic.
`naturali-score-an-agent-change`	A pass/fail verdict on an agent change.
`naturali-score-open-ended-answers`	Free-text replies graded against a rubric.
`naturali-orchestrate-several-agents`	Several agents chained into one run.
`naturali-branch-an-orchestration`	An orchestration that routes on a classification.
`naturali-pause-a-run-for-a-human-decision`	An orchestration run that waits for a person.
`naturali-model-a-process-as-a-workflow`	A process where an agent drafts and a person reviews.
`naturali-run-an-agent-on-a-schedule`	An agent that runs on a cron schedule.
`naturali-cap-project-spend`	A token budget on the whole project.
`naturali-attribute-spend-to-end-users`	Spend split per end user.
`naturali-cap-spend-per-end-user`	A token budget for each end user.
`naturali-invite-a-colleague`	A teammate added to the project by email.

Contributing

A skill draws on its tutorial steps (metadata.source), the module pages and the <Type>ResourceProperties schemas in the formations spec; when they change, change the skill. scripts/validate.mjs parses every SKILL.md front matter as YAML, as installers do, and runs in CI.

Descriptions stay under 512 characters: every installed skill's description is in every session's context.

npm ci
npm run validate

The site publishes every skill here at https://naturali.ai/.well-known/agent-skills/ on its next deploy.