These pages describe v0.19.0, the current release. That is the build the installer gives you. agentstack --version says which build you have.

Add a skill

For a skill you want across projects, put it in the central library.

Create one#

bash

$ agentstack lib new api-review

Edit api-review/SKILL.md. The frontmatter description matters because it is the small routing hint every agent sees before the full body is loaded:

---
name: api-review
description: Review API changes for compatibility, error handling, and migration risk. Use when a task adds or changes a public endpoint, request, response, or schema.
---

# Review an API change

1. Inspect the public contract and its callers.
2. Check backward compatibility and failure behavior.
3. Report concrete findings before proposing broad redesign.

Keep the body focused. Put long examples in a references/ folder and link only the ones the agent should open for a matching task.

Add it to the library#

bash

$ agentstack lib add ./api-review
$ agentstack lib add ./api-review --write
$ agentstack lib list

The first command previews the destination, content scan, and provenance. New content goes into the first writable linked source.

Import from Git in the same way:

bash

$ agentstack lib add anthropics/skills --list
$ agentstack lib add anthropics/skills --skill pdf
$ agentstack lib add anthropics/skills --skill pdf --write

Select it in a project#

toml

[toolsets.backend]
skills = ["api-review"]

Then preview, pin, and review:

bash

$ agentstack lock          # preview
$ agentstack lock --write
$ agentstack trust .

On the next trusted connection the agent sees api-review plus its one-line description. It calls agentstack_load only when a task matches. No copied project skill folder or agentstack use step is needed in the live lane.

Keep a skill inside one project instead#

Use a project-local skill when it makes sense only with that repository. The guided quick path is:

bash

$ agentstack yes

It reviews undeclared content under .agentstack/skills/, then declares, pins, and activates approved files through one human confirmation. Content from an untrusted clone is not eligible for this shortcut.

Next: Central library · Trust · Skill reference

Source of truth: docs/howto/add-a-skill.md — this page is generated from it.