Skill-inspector: Review AI Skills before Installation

A skill promising to format text may also include scripts that read files or call external services. skill-inspector helps structure the review, but a pass must include evidence and scope.

A magnifier checks a folder with components and a shield, representing skill behavior and permission reviews

skill-inspector is an inspection method, SkillSpector is a scanning tool.

The skill-inspector skill in NVIDIA/SkillSpector defines how to review other skills before installation or continued use. It combines scan results with source reading to assess whether behavior matches the stated purpose and requested permissions. Meanwhile, skillspector is the scanning CLI. Confirm it is installed and runs successfully: having a skill file does not establish CLI installation.

This article follows commit 11c8b48e of NVIDIA/SkillSpector, not a JetBrains plugin or a similarly named project. It examines commands, scripts, dependencies, and tool use: what a document skill reads, where it writes, and why it needs to send data off the machine.

An approved review still requires the agent's permission system to control files, networking, and confirmations. skill-inspector does not replace those controls with its own sandbox or behavior blocking.

Read scan results alongside a source-based review of behavior

Static scan Scan files for known risky-command, secret-access, and source/destination patterns to locate review targets. Detected text may be instructional or unexecuted code, so read context.

Semantic review reviews whether behavior fits the stated purpose. For example, a spelling skill might ask to read a token or send every file to an unexplained domain. No scan-rule match does not prove appropriate behavior; network use is not inherently dangerous when its purpose and scope are clear.

The source requires evidence to be read alongside the score. A risk score is not a percentage chance of an attack, and the review only covers the files and version inspected. When scripts or dependencies change, assess whether the earlier evidence still applies.

Prepare an inspectable copy without executing the target code

Start by identifying the skill URL or folder, retaining its commit or checksum, and reading SKILL.md alongside referenced files. Inspect scripts, dependency managers, MCP configuration, tool descriptions, and network destinations. Text-only review does not cover attached scripts.

Skills.sh lists skills from NVIDIA/SkillSpector Before installing, review source and prerequisites. For CLI use, follow the pinned README. This skill version requires notifying users when the CLI is absent and reviewing sources with lower confidence, without silently installing dependencies or runtimes.

npx skills add https://github.com/NVIDIA/SkillSpector --skill skill-inspector

This command installs the skill through Skills CLI, not the scanner. Check separately that the scanner is installed and working. This article did not run the installer or scan a package; the examples assume a working CLI.

Start scanning with LLM calls clearly separated

skillspector scan ./downloaded-skill --no-llm --format json --output report.json

The example commands are adapted from the original to use easily readable relative paths. Check the folder name before using. And don't replace the target with the entire home folder. --no-llm Disables LLM analysis for that scan, but not necessarily all network use: the README also describes OSV dependency-vulnerability lookups and remote sources

If enabling LLM analysis, check the provider, transmitted data, and cost first. The original skill instructs keeping --no-llm until the user requests LLM analysis; it must not select or switch providers itself. Check each version's README for supported or disabled providers rather than assuming support from an option's name.

If the command ends in an error, check whether the report is complete or partial. Record that static inspection is incomplete and continue reading the source. If no report exists or the scanner cannot run, state that the check failed. You cannot conclude “no risks found.”

Read evidence at its actual location, then ask whether it is necessary for the task

Hypothetical findingsQuestions requiring follow-up readingEvidence to record
Read environment variablesDoes it read only the declared service's key, or collect every value?Value names read, call sites, and forwarding
Write filesDoes it write results in the work folder, or change unrelated startup/config files?Paths, conditions, and permission requests
Make network requestsDoes the destination match the document? What content is sent?Domain, payload, and necessity
Code execution commandIs it an explanatory code block or a script that is actually invoked?Files, lines, callers, and downloaded source code

This table is hypothetical. Real findings should show rule ID, severity, path, evidence location, and why behavior is necessary or suspicious. A familiar company name does not justify downgrading unexplained HIGH or CRITICAL findings.

Record what the skill review covers and any conditions for use

Without a scanner, label the review semantic-only, do not invent scores, and state unchecked areas. Record accepted risks or baselines with reasons and evidence; do not suppress every finding to make reports look clean.

Start this process with one clearly versioned skill. Keep commands, reports, and conditions together, then use find-skills to find candidates, and review again when the skill, scripts, or dependencies change. Discovery and installation decisions are separate stages, each requiring its own evidence.

Sources

Source reviewed October 21, 2026, pinned to cited commits. This explains documentation without installing or testing the skill. Thai examples are hypothetical KrobClick scenarios. The AI-generated cover is not output from the described skill.