The Agent Skills Directory

[METADATA_POISONING]: The skill contains phrasing that strongly implies an official affiliation with Expo (e.g., "report it so Expo can improve"), despite being authored by 'hoangsoft90'. This deceptive attribution could cause users to trust the skill's instructions and dependencies under a false impression of their origin.
[UNVERIFIABLE_DEPENDENCIES_AND_REMOTE_CODE_EXECUTION]: The skill instructs the user to run npx --yes submit-expo-feedback@latest in SKILL.md. This command executes a third-party package from the NPM registry without version pinning or manual confirmation, which is particularly risky given the skill's misleading metadata.
[INDIRECT_PROMPT_INJECTION]: The skill describes a verification workflow in references/verify-on-device.md that processes external data.
Ingestion points: Reads accessibility trees and DOM structure from external web pages via agent-browser snapshot --json and simulator data via argent run describe.
Boundary markers: No explicit delimiters or instructions to ignore embedded content are provided in the verification templates.
Capability inventory: The agent has the capability to execute shell commands (npx, npm, xcrun, adb) and modify local files (migration-progress.md).
Sanitization: There is no documented logic for filtering or sanitizing external accessibility labels or metadata before comparison.
[REMOTE_CODE_EXECUTION]: The skill references and installs utilities from established sources.
Fetches the agent-browser tool from Vercel Labs' official GitHub/NPM infrastructure.
Downloads the @swmansion/argent tool from Software Mansion's official registry.

expo-web-to-native