The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains instructions that attempt to override default agent interaction patterns.
Evidence: The section ## CRITICAL: Auto-start on load explicitly commands the agent to "go straight to Step 1" and "Do not summarise the research method," which bypasses standard procedural transparency.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process a large volume of untrusted data from the live web, creating a surface for indirect instructions to influence the agent.
Ingestion points: SKILL.md (Step 2) describes navigating to and scanning feeds from Reddit, X (Twitter), and Google Search results.
Boundary markers: Absent. The instructions do not specify the use of delimiters or provide warnings to the agent to ignore instructions embedded within the retrieved post bodies or articles.
Capability inventory: The skill uses retrieved data to "Synthesise into themes" and output a Markdown table. While it does not directly pass this data to shell commands or file-write operations, the synthesis process is susceptible to manipulation by content found in social media posts.
Sanitization: Absent. There are no instructions for the agent to sanitize, filter, or validate the integrity of the text collected from external platforms before processing.

niche-research