The Agent Skills Directory

[PROMPT_INJECTION]: The skill is an orchestrator that processes untrusted external data which can influence the execution of high-privilege tools.
Ingestion points: The skill ingests task descriptions, specifications, and issue lists as signals for task classification and pipeline routing (referenced in SKILL.md and references/classification-rules.md).
Boundary markers: There are no explicit instructions or delimiters defined to separate untrusted user input from system instructions or to prevent the agent from following commands embedded in the task descriptions.
Capability inventory: The pipeline dispatches to skills that perform command execution (e.g., verify-gate running build/test tools) and persistent file system modifications (e.g., self-healing patching code, harness-updater modifying CLAUDE.md).
Sanitization: No validation or sanitization of input task data is described in the pipeline logic.
[COMMAND_EXECUTION]: The orchestrator dispatches tasks to components like verify-gate and self-healing which execute shell commands for compilation, testing, and linting, as well as automated code patching. This creates a surface for executing payloads introduced through indirect prompt injection.

skill-pipeline