The Agent Skills Directory

[METADATA_POISONING]: The skill's README and reference documentation contain fabricated historical claims, specifically asserting that Meta acquired the AI startup 'Manus AI' for $2 billion on December 29, 2025. This 'future-history' narrative is a social engineering tactic used to establish false authority and encourage the adoption of the skill's workflow.
[INDIRECT_PROMPT_INJECTION]: The skill relies on an iterative loop of reading from and writing to local files (task_plan.md, notes.md) to maintain agent state, which creates a surface for indirect prompt injection if untrusted external data (e.g., from web research or error logs) is saved to these files and subsequently interpreted as instructions by the agent. Ingestion points: notes.md for research findings and task_plan.md for error logs. Boundary markers: The provided markdown templates do not include delimiters or specific instructions to the agent to treat the stored content as untrusted data rather than active instructions. Capability inventory: The skill assumes the agent has capabilities for file system operations, shell command execution, and research tools. Sanitization: No explicit sanitization or validation steps are defined for the content being written to or read from these persistence files.

planning-with-files