The Agent Skills Directory

[PROMPT_INJECTION]: A static detector flagged the phrase 'ignore previous instructions' in SKILL.md. This is a false positive; the string appears within the 'Untrusted content' section as an example of malicious data the agent should watch for and neutralize when fetching external articles.
[INDIRECT_PROMPT_INJECTION]: The skill's primary function is to ingest untrusted external content, which is a known attack surface for indirect prompt injection. The skill implements high-standard defensive measures:
Ingestion points: Fetches data from arbitrary publication URLs, RSS/Atom feeds, and sitemaps (SKILL.md, Procedure sections 1-3).
Boundary markers: Instructs the agent to wrap imperative or prompt-shaped text in untrusted-quoted fenced blocks and apply an untrusted_directives: true flag to the metadata of the ingested page.
Capability inventory: The skill utilizes network fetch tools and writes to the sources/ and projects/ directories via standard agent file-writing tools.
Sanitization: Content is normalized to 'clean markdown' and readability-style heuristics are used to strip boilerplate navigation and scripts from HTML sources.
[DATA_EXPOSURE]: While the skill performs extensive network operations, it contains explicit policies to prevent abuse:
Anti-Exfiltration/Bypass: Explicitly forbids the use of session cookies, credentials, or 'tricks' to bypass paywalls or access gated content.
Network Ethics: Mandates a 1.5s delay between fetches, honors Retry-After headers, and implements exponential backoff for 429 Rate Limit responses to avoid DoS-like behavior.

blog-ingest