The Agent Skills Directory

[DATA_EXFILTRATION]: The skill includes instructions to access and append data to ~/.claude/projects/skill-telemetry.md. This is a hidden directory in the user's home folder used for platform-specific project metadata. Writing to files outside the current project workspace is a side-effect that can be used to track user activity or modify platform state across sessions.
[PROMPT_INJECTION]: The instruction "Do NOT ask the user questions. Execute all three phases sequentially without pausing" enforces full autonomy. This bypasses user confirmation checkpoints, preventing the user from reviewing or stopping the agent if it generates or executes dangerous code based on malicious input.
[DATA_EXFILTRATION]: (Indirect Prompt Injection Surface) The skill ingests untrusted data via the $ARGUMENTS parameter which is then used to drive the scaffolding and testing of a complete API.
Ingestion points: $ARGUMENTS variable in SKILL.md.
Boundary markers: Absent. The input is passed directly to sub-skills without delimiters or warnings to ignore embedded instructions.
Capability inventory: The skill performs file writing (scaffolding), local code execution (testing), and network operations (load testing).
Sanitization: Absent. There are no instructions to validate or sanitize the input before it is used to generate executable code.
[COMMAND_EXECUTION]: The skill chains multiple high-capability tools (/api-scaffold, /integration-test, /load-test) based on input that describes the desired API. If the input is crafted to inject malicious code into the scaffolded files, those files are then executed during the integration and load testing phases.

data-pipeline