The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill defines a workflow where human-provided labels from visual annotations are directly incorporated into AI prompts for iterative media refinement.
Ingestion points: Data enters the workflow via the anycap annotate poll command, which returns a JSON structure containing user-defined label strings from the annotation session in SKILL.md.
Boundary markers: The prompt construction logic in 'Step 3: Build a Prompt from Annotations' does not specify the use of delimiters or instructions for the AI to ignore potentially malicious commands embedded within the human labels.
Capability inventory: The skill utilizes the anycap CLI to perform network operations (uploading assets to the vendor's infrastructure), file system writes (via the -o parameter), and interactive human interaction sessions.
Sanitization: The instructions explicitly direct the agent to "Include the human's exact label text" when building prompts, which could allow an adversary (or a malicious participant in a collaborative session) to inject instructions that override agent behavior during subsequent generation steps.
[COMMAND_EXECUTION]: The skill provides numerous examples and instructions for executing shell commands using the anycap CLI tool to manage models, generate media, and handle asset deployment.

anycap-media-production