The Agent Skills Directory

[DATA_EXFILTRATION]: The skill defines processes for accessing and materializing highly sensitive personal data including personal identity profiles, employment history, housing information, and document scans (transcripts, degrees, resumes) into $HOME/.config/AgentDesk/private. The automated management of these PII-dense files presents a significant risk if the agent context or the synchronization process is compromised.
[COMMAND_EXECUTION]: The skill directs the agent to execute a specific bootstrap script (scripts/bootstrap-private-context.sh) to coordinate the cloning, decryption, and validation of private data. This instruction promotes the execution of external logic that may not be subject to the same security constraints as the primary skill instructions.
[EXTERNAL_DOWNLOADS]: The workflow relies on cloning data from a dynamic remote source defined by the AGENTDESK_PRIVATE_REMOTE environment variable. This allows the agent to fetch and process content from arbitrary remote Git repositories.
[INDIRECT_PROMPT_INJECTION]: The skill processes structured data (YAML, JSONL) and documents (PDFs) from external sources, creating a surface for indirect prompt injection.
Ingestion points: Remote storage specified by AGENTDESK_PRIVATE_REMOTE and local materialized files in AGENTDESK_PRIVATE_HOME.
Boundary markers: None mentioned; the instructions do not provide specific delimiters or warnings to ignore embedded commands within the synchronized YAML/JSONL files.
Capability inventory: File system writes, shell script execution, network operations (git clone/sync), and decryption tasks.
Sanitization: The skill mentions schema validation and readiness checks, which may catch malformed data but do not necessarily sanitize against malicious natural language instructions.

private-context-bootstrap