The Agent Skills Directory

[COMMAND_EXECUTION]: The skill orchestrates a workflow by executing several local Node.js scripts using the node command. These include scripts/verify-blind-rubric.mjs, scripts/register-cheat-score-evidence.mjs, scripts/stamp-episode-script-hash.mjs, scripts/prepare-cheat-prediction-input.mjs, and scripts/register-cheat-prediction.mjs. These scripts are not included in the skill package for verification.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from local files such as episodes/<slug>/01-口播稿.md and 00-选题.md. This content is processed by scoring and prediction tools, creating a surface for embedded instructions to influence the agent's behavior or output.
Ingestion points: Steps 2 and 4 in SKILL.md describe reading draft scripts (01-口播稿.md) and topic files (00-选题.md).
Boundary markers: The instructions do not define specific delimiters or "ignore previous instructions" markers for the ingested content.
Capability inventory: The skill has the capability to execute shell commands (Node.js scripts) and write to the local file system (e.g., lane directories, predictions/ files).
Sanitization: There is no mention of sanitization, filtering, or validation of the input markdown content before it is passed to the scoring engine.

laohan-cheat