The Agent Skills Directory

[PERSISTENCE]: The skill is designed to create institutional memory by instructing the agent to save multi-step workflows as new skills in .cursor/skills/ or .agents/skills/ and persistent constraints as new rules in .cursor/rules/. These modifications are explicitly intended to ensure knowledge and behaviors survive across different conversation sessions.
[INDIRECT_PROMPT_INJECTION]: The skill establishes a workflow where the agent automatically scans and reads all files within various workspace directories (e.g., context/, research/, docs/) at the beginning of every conversation. This behavior creates a significant surface for indirect prompt injection, as malicious instructions embedded in those files would be ingested and followed by the agent during startup.
Ingestion points: Files located in context/, .agents/, companies/, docs/, and research/ directories (SKILL.md).
Boundary markers: None identified; the agent is instructed to read relevant files without specific delimiters or instructions to ignore embedded commands.
Capability inventory: The skill has the capability to write to the file system, including creating new executable skills and system rules (SKILL.md).
Sanitization: None identified; the agent processes the content of these files as direct context for its tasks.
[DYNAMIC_EXECUTION]: The agent is encouraged to generate new instructions and logic in the form of skills and rules. Because these generated files are subsequently loaded and interpreted by the agent platform as operational logic, this constitutes a form of automated script generation and execution.
[DATA_EXFILTRATION]: The skill directs the agent to proactively aggregate research, company details, strategy decisions, and person-specific information into centralized context files. While currently focused on local storage, this consolidation of potentially sensitive information into unencrypted markdown files increases the risk of accidental exposure or easier exfiltration by other tools or skills.

saving-workspace-context