The Agent Skills Directory

[SAFE]: The skill implements robust security boundaries for social media publication. It explicitly forbids the storage of Facebook access tokens or MCP approval phrases in campaign files or asset registries. It further mandates that the agent never calls facebook_post_publish without a final preview and explicit human approval of the exact Page, copy, and media.
[SAFE]: The inclusion of a dedicated validation script (scripts/validate-post-package.mjs) provides a technical safeguard. This script scans for forbidden credential keys (e.g., access tokens, passwords, secrets) in campaign metadata before the content can move to a 'ready for preview' state.
[PROMPT_INJECTION]: A static detector flagged references/higgsfield-routing.md for a possible DAN/jailbreak attempt. Upon manual review, this is a false positive. The flagged text consists of standard negative constraints (e.g., 'Do not bypass its backend prompt enhancer', 'Never silently fall back to an older model') which are intended to preserve production quality and safety, not to override agent filters.
[INDIRECT_PROMPT_INJECTION]: The skill exhibits a standard vulnerability surface as it processes external research data and user-supplied content which is then interpolated into prompts for media generation and social posting.
Ingestion points: Research data stored in research.md, brand profiles in brand-profiles/*.yaml, and campaign source material.
Boundary markers: Phase 7 and Phase 8 require explicit user approval of a 'post-package' and a platform-rendered preview before any publication occurs.
Capability inventory: The skill can perform network operations via facebook_post_publish and trigger remote media generation via Higgsfield tools.
Sanitization: The skill uses a JSON-based packaging format and a validation script to ensure data conforms to expected schemas and does not contain unauthorized instruction fields.

facebook-content-studio