AGENT SKILLS · PRE-INSTALL REVIEW

A valid skill can still be dangerous.

Paste SKILL.md and review two separate questions: does it follow a portable structure, and does its content request risky tools, hidden authority, credential access, or remote execution?

FrontmatterPrompt injectionTool scopeLocal-only

STATIC SKILL AUDIT

SKILL.md scanner

SKILL.md0 lines · ~0 tokens

Checks portable frontmatter, activation quality, tool permissions, suspicious instructions, secret patterns, remote URLs, shell execution, and context size.

Structure

Required delimiters, name and description, portable naming, unknown fields, and instruction body.

Authority

Hidden overrides, attempts to suppress approval, requests for credentials, and instructions to conceal actions.

Execution

Shell downloads, privilege escalation, destructive commands, tool scope, and untrusted remote dependencies.

Context quality

Activation phrases, actionable steps, references, excessive length, and claims the scanner cannot verify.

TRUST BOUNDARY

Skills are instructions plus supply chain.

A skill can influence when an agent activates, which tools it uses, what files it reads, and where it sends data. Static text review catches visible warning signs, but installation trust also depends on the publisher, referenced files, scripts, version history, and the behavior of every allowed tool.

Scan before install

Read the entire selected skill
Inspect scripts and referenced resources
Review allowed tools against the task
Search for network destinations and secrets
Pin the source and record provenance

Recheck after updates

Diff instructions and frontmatter
Audit new scripts and binary assets
Review expanded tool permissions
Retest activation boundaries
Remove abandoned or duplicated skills

SKILL.md validatorPortable frontmatter and structure →Agent Skill security guideThreat model and review checklist →Skill description guideWrite reliable activation text →