The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to fetch live content from a Wikipedia page ("en.wikipedia.org") and treat it as an authoritative source of truth for its writing and auditing rules. Since Wikipedia is a community-edited platform, an attacker could modify the page to include malicious instructions that the agent might ingest and follow when performing its tasks.
Ingestion points: SKILL.md (Source of Truth section) instructs the agent to fetch raw wikitext from "https://en.wikipedia.org/w/index.php?title=Wikipedia:Signs_of_AI_writing&action=raw".
Boundary markers: No explicit markers or "ignore embedded instructions" warnings are used when fetching the live Wikipedia content.
Capability inventory: The skill has access to shell commands like curl and grep, and it has the capability to read and write files to maintain its local snapshot of the guidelines.
Sanitization: There is no evidence of sanitization or validation of the remote content before it is processed by the agent.
[EXTERNAL_DOWNLOADS]: The skill uses curl to fetch data from Wikipedia's API and raw wikitext interface to update its internal snapshots and perform thorough audits.
[COMMAND_EXECUTION]: The skill utilizes shell commands for its core workflow. It uses curl for remote data retrieval and grep with complex regular expressions to perform mechanical self-reviews on text drafts to detect AI-generated artifacts.

natural-writing