The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill instructs the agent to ingest and evaluate external evidence such as measurements, recordings, and renders. This data-processing workflow creates an attack surface where instructions embedded within these evidence files could influence the agent's judgment or behavior.
Ingestion points: Identified in SKILL.md (Step 3: Score honestly from real evidence) and templates/critic-prompt.md (Evidence parameter).
Boundary markers: The templates do not utilize explicit delimiters or markers to isolate the external evidence content from the judge's logic.
Capability inventory: The skill is limited to file reading and text generation. No network exfiltration, arbitrary command execution, or sensitive credential access capabilities are identified.
Sanitization: No sanitization or validation of the content of evidence files is specified.
[SAFE]: The use of an independent subagent judge ('Independent critic prompt') is a robust design pattern that isolates the evaluation logic from the task execution logic, which mitigates the impact of instructions intended to bias the scoring process.
[SAFE]: All identified file operations and references (e.g., saving to qa/ directories) are local to the project context and align with the stated purpose of documentation and QA reporting.

workflow-score-to-target