The Agent Skills Directory

[COMMAND_EXECUTION]: The skill relies on the execution of shell commands via the codex exec and higgsfield CLI tools to generate and manipulate images. It also references a local execution script at tools/fal_edit.py for image editing tasks.
[PROMPT_INJECTION]: Instructions in Part 2 suggest using CLI tools specifically to circumvent content filters present in web applications (e.g., blocking currency notes as "protected content"), which constitutes a safety guideline bypass.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data (briefs and copy) from the ad-strategy agent. This data is interpolated into prompts used for image generation and command execution.
Ingestion points: Reads campaign briefs and copy from the ad-strategy agent.
Boundary markers: Utilizes "Shared Constraints" blocks and explicit hard-negatives to attempt to restrict the image model's behavior.
Capability inventory: Includes shell command execution (codex, higgsfield), file system writing (workspace-write), and local script execution.
Sanitization: Implements a vision-based gate through the static-qa agent to verify outputs before human review, although this does not prevent injection at the input stage.
[DYNAMIC_EXECUTION]: The skill describes the programmatic generation of graphical elements (such as security marks and stars) using code libraries like Pillow or SVG rendering to overcome limitations in generative models.

static-ad