The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data by using the Firecrawl toolkit to extract branding information and screenshots from third-party websites. While this creates a surface for indirect prompt injection (where malicious content on a scanned site could attempt to influence the agent's output), the skill effectively mitigates this risk by requiring that ad copy be grounded in the source material and mandating explicit human review before any campaigns are activated.
[Ingestion points]: Data enters the agent's context through firecrawl_branding_extract and firecrawl_screenshots_create (referenced in references/discovery.md).
[Boundary markers]: The instructions in SKILL.md explicitly direct the agent to 'ground ad copy in what the page actually says' and warn 'Never invent copy for a page you have not read.'
[Capability inventory]: The skill utilizes the meta_ads_campaign_create, meta_ads_ad_create, and meta_ads_campaigns_activate tools to interact with the Meta advertising platform.
[Sanitization]: The workflow relies on mandatory human-in-the-loop confirmation ('Activate only when the user approves') to verify all generated content and targeting before it goes live.
[DYNAMIC_EXECUTION]: The skill includes functionality for building performance dashboards via data_apps_build, which involves generating Python script snippets (prefab_python) for the dashboard interface. This use case is standard for data visualization within this ecosystem and includes specific security guidance to ensure that Meta API credentials are not injected into the dashboard runtime, adhering to safe implementation practices.

meta-ads