The Agent Skills Directory

[DATA_EXFILTRATION]: The skill includes diagnostic examples that demonstrate how to access sensitive system data for troubleshooting purposes. In SKILL.md, it suggests commands like 'env | grep IDENTITY' to inspect environment variables and uses the macOS 'security' utility to list keychains and signing identities. This pattern involves exposing potentially sensitive configuration and credential storage information.
[COMMAND_EXECUTION]: The skill provides a shell script, 'find-polluter.sh', which executes local project tests using 'npm test'. The instructions also describe the use of various shell commands such as 'git init', 'codesign', and 'security' to gather evidence during the debugging process.
[PROMPT_INJECTION]: The skill uses extremely strong directives and imperative language (e.g., 'ALWAYS find root cause', 'NO FIXES WITHOUT ROOT CAUSE') to override default agent behavior and prevent the use of shortcuts. This structure acts as a behavioral constraint to enforce a specific methodology.
[PROMPT_INJECTION]: The skill creates a surface for indirect prompt injection by directing the agent to ingest and analyze untrusted external data like error logs and stack traces. Ingestion points: Error messages, logs, and test outputs in SKILL.md and root-cause-tracing.md. Boundary markers: None provided to isolate data from instructions. Capability inventory: Execution of shell commands and project tests. Sanitization: No validation or sanitization of ingested data is specified before the agent uses it to inform command execution.

systematic-debugging