Skills are reusable playbooks that teach Stella how to perform a task. A Skill is a directory with a SKILL.md file and may include reference files or scripts. Stella reads the files selected for the active Agent at the start of a turn.

Stella reads four typed resource roots:

Project Skills under .agents/skills/ are considered before package Skills. For each package name, Stella selects the most specific complete package from the four typed roots in this order:

user_agent > user > system_agent > system

A selected package replaces the complete package at broader scopes. Its files, Skills, CLI entries, environment bindings, and MCP declarations are resolved as one unit. An independent copy has no parent link and does not receive later updates from the source. A standalone Skill follows the same scope precedence by its name.

A settings.json file can disable a resource or add administrator-only forbidden and disabled_tools entries. A disabled or malformed winner masks an inherited resource. A narrower personal setting cannot override an administrator prohibition.

Stella captures the resource files at turn admission. An edit made during a turn takes effect on the next turn; reads in the current turn keep the captured content. The capture is also the view used by search, prompt loading, CLI preparation, and MCP discovery.

Choose the destination before every install or upload. Stella does not infer a scope from the conversation. The Web UI and API can create, copy, edit, and delete a complete package or standalone Skill in a writable scope. Package edits replace the complete package tree. File edits use the current content digest when the API provides one, so a stale edit is reported as a conflict. Direct writes from a shell are supported, but they do not provide a multi-file transaction or a filesystem compare-and-swap guarantee.

An uploaded archive must contain one Skill directory with SKILL.md. A copied package is a new resource under the destination owner. It keeps its own files, settings, credentials, and OAuth grants; editing the source later does not upgrade the copy.

Ask Stella to find an installed Skill, then load the one that matches the task. skill_installed_search searches Skills visible to the active Agent; it does not search a marketplace. skill_load reads the selected files into the current session's temporary sandbox directory. The returned path is disposable and is the path to use for the Skill's scripts and bundled files.

A failed preparation keeps the selected package masked for that turn. Stella does not silently revive a lower-priority same-name resource. Native tools are separate system capabilities and are not granted by a Skill or package with a matching name.

To update a Skill or package, edit or replace its complete files in the intended scope. Read the current resource before retrying after a digest conflict. A successful write affects the next turn. The current turn finishes with its captured Skill and package view.

Disabling a package or Skill blocks it from future selection. It does not revoke an OAuth grant or remove resource bytes. Deleting an MCP declaration file removes that declaration only. Use the MCP Disconnect action to revoke the matching locally stored grant, close its connection, and block late callbacks or refreshes; remote provider revocation is not guaranteed. The two operations are independent.

Uninstalling a Skill or deleting a package removes it from future selection. Stella keeps historical usage and changelog evidence needed by Skill reflection. It does not run an online garbage collector for resource bytes or caches. Cleanup requires an explicit maintenance action after the owning process and its descendants are proven stopped.

The local backend enforces its configured sandbox policy but cannot prove that detached descendants stopped. The none backend provides no reliable process isolation. Do not treat a normal turn close as proof that descendants stopped, and do not clear retained resource bytes using a time-to-live or a guessed process ID. The none backend is unsuitable when cleanup requires a hard process boundary.

---
name: my-deploy-script
description: Deploy the application to production.
---

# Deploy to production

1. Run the test suite.
2. Build the production bundle.
3. Verify the deployment is healthy.

name must use lowercase letters, numbers, and hyphens, and be at most 64 characters. description is required and appears in search results. disable-model-invocation is optional. Set it to true when Stella should invoke the Skill only after an explicit request; it does not disable the Skill or change edit permissions.

For team workflows, commit a Project Skill under .agents/skills/. For a shared or personal Skill, create or upload the complete directory in the intended scope.

Search before creating a new Skill. Keep one Skill focused on one workflow. Load the Skill after creating or updating it and run a small representative task before relying on it.