The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill's core workflow involves an interactive discovery session where it ingests untrusted user input (process names, pain points, constraints) and interpolates this data into the system prompts of a generated multi-agent team.
Ingestion points: User input collected during the discovery phase described in SKILL.md.
Boundary markers: Absent. The skill does not specify the use of delimiters or instructions to the LLM to ignore potentially malicious directions embedded in the user's business process description.
Capability inventory: The skill utilizes Bash, Write, and Glob tools to create configuration files and directory structures. Furthermore, the templates in references/team-templates.md frequently assign the Bash tool to the generated agents, providing a significant capability surface for any agent influenced by injected instructions.
Sanitization: The skill includes a rule to avoid hardcoding secrets, but it lacks specific sanitization or filtering logic to prevent 'jailbreak' or 'override' instructions from being passed from the discovery phase into the generated prompt.md files.
[COMMAND_EXECUTION]: The skill uses the Bash tool for its operations and explicitly templates Bash access for generated agents (e.g., Sales Coordinator, Lead Qualifier, Technical Support Engineer). While these tools are necessary for the skill's stated purpose of automating business workflows, their inclusion in the generated configuration increases the potential impact of a successful prompt injection attack on the resulting agent team.

agent-team-builder