Back to blog

Install an agent skill with the SourceWeft CLI, and check that the files match what was scanned

A step-by-step tutorial: search for a skill, read its details, install it for Claude Code, Codex or Cursor, and see how the CLI checks every file against the hashes recorded at scan time. Real commands and output.

SOSourceWeftNov 20, 20266 min read

Last updated: 2026-11-20

What you will do

You will search the SourceWeft skills directory from a terminal, read a skill's details before installing, install it for a local agent, and then check that what is on disk is what the directory scanned.

Every command below was run with the published CLI, version 0.3.0, in an empty folder on 2026-11-20. You need Node.js and a terminal. You do not need a SourceWeft account.

If you want to understand how to judge whether a skill is appropriate for you, read How to check an agent skill before you install it first. This tutorial is about the mechanics.

What "verified" means here

The directory records, for each skill, the source repository, the exact commit it was scanned at, and a SHA-256 hash for every file. The CLI downloads the files from that repository at that commit and compares each one to the recorded hash. If anything differs, it installs nothing.

That check proves one thing: the files you install are the files that were scanned. It does not prove the files are harmless. For that, you read the skill and its scan notes. The two checks do different jobs.

npx @sourceweft/cli skills search "pdf forms"

Each line shows the install name, a scripts marker if the skill ships scripts, and a short description. A few of the results we got:

@github/pdftk-server                                       Skill for using the command-line tool pdftk (PDFtk Server) for workin…
@claude-office-skills/pdf-form-filler-0a843e12             Fill out PDF forms programmatically and extract form data
@jwynia/pdf-generator                             scripts  Create and manipulate PDF files programmatically. Use when the user n…

The install name, such as @github/pdftk-server, is what you pass to the other commands.

Step 2: Read the details before you install

npx @sourceweft/cli skills info @google/gke-kueue-troubleshooting

This is the output for a skill that ships scripts and carries scan flags:

@google/gke-kueue-troubleshooting  (Gke Kueue Troubleshooting)
Diagnose a GKE batch or AI workload (…

Version:    e96f2070b42c11e184e662cfe1fda8b20c9e2140
License:    not stated
Source:     https://github.com/google/skills/tree/e96f2070b42c11e184e662cfe1fda8b20c9e2140/skills/cloud/gke-kueue-troubleshooting
Trust:      Featured — a featured publisher, not reviewed by a SourceWeft admin (files are still hash-checked)
Files:      24
Scripts:    yes — this skill includes executable files
Scan flags: egress:fetch, script:out-of-bundle-path, secrets:env-access

Read these lines in this order:

  • Source is the repository and commit. The commit is the version you will get.
  • Trust says whether a SourceWeft admin reviewed the skill. "Featured" and "Community" mean no admin did.
  • Scripts tells you whether anything in the skill is code. If so, it can run on your machine.
  • Scan flags are the automated scan's notes. What they mean is explained in the article linked above.

Step 3: Install

npx @sourceweft/cli skills install @google/gke-kueue-troubleshooting

The CLI prints the same details and asks you to confirm. If the skill ships scripts, the question is "This skill includes scripts your agent may run. Install it?"

In a terminal that is not interactive (a script or a CI job), the CLI will not guess. It stops with:

Error: Confirmation required. Re-run with --yes to proceed without asking.

and exits with a non-zero status, having written nothing. Use --yes only when you have already read the details.

By default the skill is installed for Claude Code, into ~/.claude/skills/<name>/. To choose another agent, use --agent:

npx @sourceweft/cli skills install @github/pdftk-server --agent codex

npx @sourceweft/cli skills agents prints the full list with the folder each agent uses. For example, claude-code uses ~/.claude/skills, cursor uses ~/.cursor/skills, and codex uses ~/.agents/skills. Add --scope project to install into the current project's folder instead of your home directory.

To try an install without touching your real agent folders, point it at an empty folder:

npx @sourceweft/cli skills install @google/gke-kueue-troubleshooting --dir ./out --yes

We ran this, and it wrote 24 files into ./out/gke-kueue-troubleshooting/: the skill's own files plus one file the CLI adds, .sourceweft.json.

Step 4: See what was recorded

Open .sourceweft.json. It holds the registry address, the skill's install name, the version, the source repository and commit, the license, and a hash for every file:

{
  "installRef": "@google/gke-kueue-troubleshooting",
  "version": "e96f2070b42c11e184e662cfe1fda8b20c9e2140",
  "source": {
    "repoUrl": "https://github.com/google/skills",
    "commitSha": "e96f2070b42c11e184e662cfe1fda8b20c9e2140",
    "subpath": "skills/cloud/gke-kueue-troubleshooting"
  },
  "files": { "references/attribution.md": "b419a7d3…", "…": "…" }
}

We recomputed the SHA-256 of all 24 files with a separate script, and every one matched the recorded value. You can do the same with shasum -a 256 and compare.

Step 5: Check the install later

npx @sourceweft/cli skills doctor

doctor compares what is on disk with what was recorded. Right after the install above, with --dir ./out:

OK — 1 installed skill(s), no problems found.

We then appended one line to scripts/snapshot.py and ran it again:

note     …/gke-kueue-troubleshooting: 'scripts/snapshot.py' was edited

doctor reports edited files as a note, missing files as a problem, and extra files as a note. It tells you that something changed. It cannot tell you whether the change was yours.

Two more commands behave the same way around local edits:

  • skills list shows the install with the state modified.
  • Running install again over a modified folder stops with an error and asks for --force, rather than overwriting your changes. With --force it restores the recorded files and prints Updated ….

Step 6: Update and remove

npx @sourceweft/cli skills update --dry-run
npx @sourceweft/cli skills update
npx @sourceweft/cli skills remove @google/gke-kueue-troubleshooting

--dry-run shows what an update would do without doing it. A new version can bring new scripts, so read the new version's details the same way you read the first.

Without the CLI: ask your agent to do it

Each skill page has an Agent tab with a prompt you can paste into your agent. It tells the agent to read the skill's install.md, show you the source, license, and whether it ships scripts, wait for your OK, and ask before changing anything else on your machine. Use this if you would rather not run commands yourself.

What this does not do

  • It does not review the skill for you. A file that matches its hash can still be a bad idea to run.
  • It does not protect you from a skill that was already problematic when it was scanned. That is what reading the skill and its scan notes is for.
  • It does not watch the upstream repository for changes after you install. skills update is how you find out about a new version.

FAQ

How do I install an agent skill from the command line? Run npx @sourceweft/cli skills install <name>. It prints the skill's source, license, and scan notes, asks you to confirm, and writes the files for your agent. By default it installs for Claude Code.

Which agents does the CLI support? npx @sourceweft/cli skills agents lists them, with each agent's skills folder. The list includes Claude Code, Codex, Cursor, Gemini CLI, Windsurf, and GitHub Copilot, among others.

What does the CLI check before it writes a file? It downloads the skill from its source repository at the recorded commit and compares every file to the SHA-256 hash recorded when the skill was scanned. If the download does not match, it installs nothing.

Do I need a SourceWeft account? No. The commands in this tutorial ran without signing in.

Can I try an install without changing my agent's folders? Yes. Use --dir <folder> to install into an empty folder, and skills remove or delete the folder when you are done.

What if I edit an installed skill? skills doctor and skills list will show that it was edited. Installing over it requires --force, so your changes are not overwritten by accident.

Shareable URL

Keep reading

Related essays

All posts

Agent skills for PDF, Word, Excel and PowerPoint work: what is available and how to pick

Which agent skills handle PDF, Word, Excel and PowerPoint files, what each ships (files, scripts, scan notes, license), and how to install one for Claude Code with a command. Checked against the live directory on 21 November 2026.

Read next

What are Claude skills, and how do you find and install them?

A plain explanation of Claude skills (agent skills): what a skill folder contains, where to find skills, and two ways to install one, with a command or by hand. Includes what to check before installing.

Read next

How to install skills and MCP servers in Claude Code, from an empty folder

Start from an empty folder and install a skill and an MCP server for Claude Code: search, inspect before installing, where the config files go, and how to check and remove them. Every command was actually run.

Read next