The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes data from various project artifacts to build a solution package.
Ingestion points: Reads multiple Markdown files including method explanations (qx_final_method_explanation.md), result analyses (qx_final_result_analysis.md), and decision logs (qx_decision_log.md).
Boundary markers: The skill uses structured section mapping and explicit transcribing rules (e.g., using decision_id for provenance) to maintain data boundaries.
Capability inventory: Performs local file reads and writes to generate Markdown summaries and a JSON data-freeze file (frozen_numbers.json). It does not have network access or arbitrary command execution capabilities.
Sanitization: Implements a mandatory human-in-the-loop sign-off gate (Gate G4.5) where a human must review and decide on confidence levels and claim scopes before the numerical freeze is emitted.
[DATA_EXFILTRATION]: No network operations or unauthorized data transmission patterns were identified. Access is restricted to project-local paths within the methods/, results/, and robustness/ directories.
[COMMAND_EXECUTION]: No shell command execution or dynamic code evaluation patterns were detected. The skill focuses on text extraction and organization.
[PROMPT_INJECTION]: The skill contains internal process labels like 'CRITICAL' and 'IMPORTANT', but these are used in a functional context to enforce workflow gates and do not attempt to bypass agent safety guidelines.

solution-package-builder