Skip to content

Guide · Agent skill

How to use the Skill Installer agent skill (skill-installer)

Updated · written with AI from the project's own documentation and checked against our catalogue data

The Skill Installer is an agent skill that acts as a gatekeeper when you want to add a third-party skill from a registry. Instead of copying files in blindly, it checks the source, shows you the raw instructions, runs trust checks and a scan, and waits for your explicit yes before writing anything to disk. This guide explains what agent skills are, how to install this one, how to trigger it, and where its protections stop.

Skill Installeranthropics/claude-for-legal · Apache-2.0

Quick start

  1. 1Download and unzip. Download the skill ZIP from this page and unzip it into ~/.claude/skills/ or a project's .claude/skills/ folder.
  2. 2Or upload in the Claude apps. Go to Settings → Capabilities → Skills and upload the skill.
  3. 3Optionally create an allowlist. Add allowlist.yaml under the claude-for-legal plugin config and choose permissive or restrictive mode.
  4. 4Ask for an install. Say "install [skill name]" or give a direct skill URL.
  5. 5Review the raw SKILL.md. Read the full raw file and any flagged injection patterns shown above it.
  6. 6Check the scan results. Review the trust-check findings and the skills-qa verdict.
  7. 7Approve explicitly. Type yes only if you want the files written. Anything else leaves the disk untouched.

What an agent skill is and what the Skill Installer does

An agent skill is a folder containing a SKILL.md file of instructions, plus optional references, templates, commands and agents that Claude can follow when a task matches the skill's description.

The Skill Installer makes Claude follow a fixed workflow when you ask to add a community skill. It reads the allowlist first, fetches the candidate skill, displays its full raw SKILL.md, runs structural trust checks (hooks, MCP servers, tool permissions, file writes, network calls), and runs a skills-qa scan. Only after you type an explicit yes does it copy the skill into place.

  • Reads an allowlist before fetching anything
  • Shows the raw SKILL.md in full, not a summary
  • Checks hooks, MCP servers, tool permissions, file-write targets and network calls
  • Runs skills-qa and reports the verdict and heuristic findings
  • Writes nothing to disk until you approve

Install the Skill Installer skill

Download the skill ZIP from this page and unzip it into one of two locations. Use ~/.claude/skills/ to make it available across your projects, or a project's .claude/skills/ folder to limit it to that project.

If you use the Claude apps instead, upload the skill under Settings → Capabilities → Skills.

  • Personal: unzip into ~/.claude/skills/
  • Project only: unzip into the project's .claude/skills/ folder
  • Claude apps: Settings → Capabilities → Skills, then upload

Set up an allowlist (optional)

The installer reads allowlist.yaml from the claude-for-legal plugin config at ~/.claude/plugins/config/claude-for-legal/legal-builder-hub/allowlist.yaml. The file sets a mode (permissive or restrictive), a list of trusted registries and publishers, and the licences and MCP connectors allowed.

If the file does not exist, the installer runs in permissive mode with empty lists. It then warns about unknown sources but does not refuse them. The source notes that running /legal-builder-hub:cold-start-interview can create the file.

  • mode: permissive warns on unknown sources and continues
  • mode: restrictive refuses sources not on the allowlist
  • Without an allowlist, nothing is refused, only flagged
mode: permissive

Trigger the skill with natural requests

The skill's description tells Claude to use it when you say things like "install [skill]", when you pick install from browse, or when you give a direct skill URL. You can also invoke it as /skill-installer with a skill name or registry URL.

Be specific about the skill and where it comes from so the allowlist check has something to check against.

  • "Install the contract-redlines skill from the registry"
  • "Install this skill: https://github.com/example/skills/tree/main/review-skill"
  • "/skill-installer [skill name or registry URL]"
/skill-installer [skill name or registry URL]

Review the raw SKILL.md and approve or decline

After fetching, the installer shows the raw SKILL.md above any findings. Injection patterns such as attempts to override instructions, external URLs, hidden unicode or out-of-scope file writes are flagged above the raw content. Read it yourself rather than relying only on the summary.

The approval prompt asks: Proceed? (yes / no / show full). Only a fresh yes typed by you allows the install. Earlier messages do not count as approval.

  • Read the raw content, not just the summary
  • Check the trust-check and skills-qa findings
  • Answer yes only when you are ready for files to be written
  • Choose show full to see everything before deciding

Safety limits to understand

The skill is a set of instructions to Claude, so it cannot fully secure itself against prompt injection. A clever injection inside a third-party SKILL.md could try to make Claude skip the raw display, report a clean scan, or write files early. The protections reduce that risk but do not eliminate it.

For the strongest guarantee, the skill recommends running the fetch and analysis in a read-only subagent with Read and WebFetch only. The skill runs no scripts itself, but it does use the network to fetch candidate skills.

  • Runs scripts: no
  • Uses network: yes
  • Analysis tools: Read, WebFetch, Glob; Write and Bash only at the install step
  • Requires skills-qa for the scan step
  • Restrictive mode refuses unknown licences and sources; permissive mode flags them

What people use it for

  • Reviewing a community contract-clause skill before a firm lets lawyers run it locally
  • Checking which hooks, MCP servers and file-write targets a developer's new skill declares before installing it
  • Reading the raw instructions of a skill from a public registry before deciding whether to keep it
  • Enforcing a restrictive allowlist so only approved registries and publishers can install skills

Frequently asked questions

Is the Skill Installer free to use?

The source does not state pricing or licence terms for your use beyond the item's listed licence, which is Apache-2.0. Check the licence file in the repository for the terms that apply to you.

What do I need to run the Skill Installer?

It needs Read, WebFetch and Glob for analysis, and Write and Bash only at the install step. It uses skills-qa for the scan step. An allowlist.yaml file is optional; without one the installer runs in permissive mode.

Where do I put the skill so Claude can find it?

Unzip it into ~/.claude/skills/ for all projects or into a project's .claude/skills/ folder for one project. In the Claude apps, upload it under Settings → Capabilities → Skills.

Does the installer write files without asking?

No. The skill says it writes nothing to disk until you type an explicit yes. Its workflow requires a fresh approval, and earlier messages do not count.

Can the Skill Installer guarantee a skill is safe?

No. The skill states that a well-crafted prompt injection inside a third-party skill could manipulate the AI, so its protections reduce rather than eliminate that risk. Reading the raw SKILL.md yourself is the main check you control.

What happens if I have no allowlist?

The installer proceeds in permissive mode with empty lists. It warns about unknown registries and publishers but does not refuse them, so it is not suitable if you need a hard block.

New agent skills, repos and AI tools — every week

The best new skills and open-source projects, the fastest-growing repos, and one free prompt. Five minutes to read. Unsubscribe anytime.

Our verdict

Use the Skill Installer if you want a visible review step before community skills reach your machine, especially in a legal team or firm that wants to read raw instructions first. Pick something else if you want instant installs with no manual approval, or if you need strong security guarantees, since the skill itself says AI-mediated trust cannot be fully secured against prompt injection.

Download the skill