# Agent Skills: review before installation

An Agent Skill packages instructions and supporting files for a repeatable task. Review what it can read, run and send before adding it to an agent; a useful description is not evidence that the code is safe.

## What a skill contains

The [Agent Skills format](https://agentskills.io/) uses a SKILL.md file with a name, description and instructions, plus optional scripts, references and assets. Those files can change the agent's behavior. Read the full package, not just the summary or a popularity score.

OwnAgents uses skills as part of a scoped workflow discussion. It does not offer an unrestricted marketplace where every published skill is approved for customer use.

## Review access and side effects

- Identify the real task, expected inputs and output.
- Inspect bundled scripts, dependencies and any installation commands.
- Check which files and accounts may be read or changed.
- List network destinations and information sent to each one.
- Check how credentials are supplied and whether they could appear in logs.
- Identify sending, publishing, deletion, purchases and other external changes that need approval.

Reject instructions that ask the agent to ignore its operating policy, disclose secrets, or treat a web page as an authority to act. A skill does not gain permission simply because it is installed. Prefer extending an already reviewed workflow when a new package adds no useful capability.

## Test in an isolated environment

Use synthetic or non-sensitive input first. Compare the actual output with the requested task, inspect the files and network requests it produces, and check failure behavior. Test whether it stops when required access or approval is missing. A reassuring model response is not a substitute for observing those effects.

## Keep a review record

Record the source and reviewed version, supported inputs, required tools, data access, output location and approval conditions. Name the verification steps and a way to disable the skill or undo supported changes. Review meaningful updates before replacing the tested version; the earlier review does not automatically cover new files or permissions.

These checks reduce avoidable risks but do not certify a package as safe for every environment. Credentials and confidential customer records do not belong in the public questionnaire or a shared example.

## Add skills after defining the workflow

Read the [deployment readiness checklist](https://ownagents.net/agent-deployment/) for environment and permission decisions. The [Personal AX service overview](https://ownagents.net/ai-agent-services/) explains the consultation scope. [Start with one task](https://ownagents.net/#starter) rather than choosing a catalog of automations before deciding what you need.