The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes user-provided content, such as scripts, character descriptions, and reference images, which it then incorporates into prompts for downstream video generation models. This creates a surface for indirect prompt injection.
Ingestion points: The skill ingests untrusted user data in SKILL.md (Decision Router, Prompt Contract) and references/asset-locked-workflow.md (Asset Rules, Clip Plan).
Boundary markers: While no explicit character delimiters are specified for user input, the skill uses a node-based architecture to isolate reference data from model instructions.
Capability inventory: The skill has access to tools for creating workflows (workflow_author), saving versions (save_workflow_version), and launching generation runs (launch_workflow_run and the postplus CLI).
Sanitization: The skill implements a robust mitigation strategy through a three-tier quality gate system (L0 Hard Check, L1 Self Review, L2 Deep Review) and requires explicit human approval for all cost-incurring or persistence actions.
[COMMAND_EXECUTION]: The skill utilizes the postplus command-line interface to manage workflows. These commands are legitimate tools provided by the vendor (postplusai) for platform management. The skill correctly identifies that these operations require a logged-in session and enforces the use of safety flags (e.g., --confirm) for sensitive operations like launching a workflow run.

workflow-creation