The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill audits untrusted artifacts which serves as a primary ingestion point for potentially malicious instructions hidden in data.
Ingestion points: Untrusted content is read into an artifact.md file and subsequently processed by multiple critic and judge subagents.
Boundary markers: The skill implements robust defenses, including mandatory 'CONTENT ISOLATION' warnings and 'UNTRUSTED DATA' delimiters in all subagent prompt templates to prevent the models from obeying instructions within the artifacts.
Capability inventory: The skill utilizes Bash, git, the Task tool for spawning subagents, mcp__pal__chat for external provider calls, and Write for file management.
Sanitization: The skill employs a deterministic counting-substrate hierarchy using jq, grep, and wc to verify data instead of relying on LLM estimation, and uses a rationalization auditor to ensure final reports accurately reflect independent judge verdicts.
[COMMAND_EXECUTION]: The agent uses the Bash tool to perform repository-wide searches with grep, analyze changes via git, and run a local utility script aggregate_ensemble_judges.py to perform deterministic vote aggregation across the model panel.
[DYNAMIC_EXECUTION]: The workflow involves spawning parallel subagents using the Task tool and making dynamic calls to external LLM providers (OpenAI, Google) via mcp__pal__chat. These tasks involve assembling prompts that incorporate data extracted from the analyzed artifacts at runtime.

deep-qa-ensemble-v1