The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes content from external websites (Product Hunt, LinkedIn, Reddit, etc.) via the agent-browser tool to generate launch strategies, creating a surface for indirect prompt injection.
Ingestion points: Untrusted data enters the agent context during 'Research Mode' via browser automation commands documented in references/launch-type-assessment.md.
Boundary markers: The skill implements strict confirmation gates in SKILL.md (Response Protocol steps 6, 8, and 9) requiring user approval before drafting or saving content.
Capability inventory: Capabilities include browser/network access, local file system writes for deliverables, and shell command execution for research and discovery scripts.
Sanitization: The skill relies on mandatory user review of all generated drafts and saved files as the primary mitigation strategy.
[COMMAND_EXECUTION]: The skill provides templates for executing shell commands in references/launch-type-assessment.md and references/shared-patterns.md for tool discovery and browser research. These commands are tailored to the skill's primary marketing research function and do not exhibit privilege escalation patterns.
[EXTERNAL_DOWNLOADS]: The skill recommends installing agent-browser from the trusted vercel-labs GitHub organization. This is a functional dependency for competitive intelligence and is documented neutrally as a configuration step.
[DYNAMIC_EXECUTION]: The skill executes a vendor-provided tool discovery script (tool-discovery.sh) to validate the execution environment. This script is part of the broader pawbytes marketing suite infrastructure.

paw-mkt-launch