The Agent Skills Directory

[PROMPT_INJECTION]: The skill uses authoritative language such as "single mandatory dispatcher," "invoke it first, unconditionally," and "Always invoke project-orchestrator first" to override the agent's natural selection of tools and skills. These instructions attempt to hijack the agent's reasoning process and prioritize the developer's custom workflow over default agent behavior.
[INDIRECT_PROMPT_INJECTION]: The skill defines a process for ingesting untrusted data from the repository and passing it to subagents, which could allow external data to influence the agent's instructions.
Ingestion points: The skill reads project-specific configuration from .claude/dev-conventions.json and gathers codebase context through the codebase-domain-map tool.
Boundary markers: There are no instructions for using delimiters or boundary markers when interpolating this external data into subagent prompts.
Capability inventory: The skill is capable of writing to the local filesystem (creating/updating .claude/dev-conventions.json) and delegating tasks to subagents with injected context.
Sanitization: No sanitization or validation steps are defined for the data read from configuration files or codebase indices before it is used to generate prompts for other agents.

project-orchestrator