The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests project descriptions through the $ARGUMENTS variable and reads project context from CLAUDE.md and README.md using a sub-agent.
Ingestion points: $ARGUMENTS in SKILL.md; project files like CLAUDE.md and README.md via the Explore agent in Step 0.1.
Boundary markers: The skill implements a 'Scope Gate' (Step 0.2) that validates the project description. If the input is too vague (e.g., 'make things better'), the agent is instructed to stop and ask clarifying questions instead of proceeding.
Capability inventory: The skill utilizes TaskCreate, TaskUpdate, and Write tools to manage tasks and output visualizations.
Sanitization: The agent is instructed to perform a structured analysis phase and uses a scope gate to ensure the task breakdown is based on concrete deliverables rather than untrusted, ambiguous input.
[EXTERNAL_DOWNLOADS]: The WebFetch tool is listed in the allowed-tools section of the frontmatter. While this gives the agent the capability to access external URLs, the instructions do not contain any hardcoded malicious URLs or patterns that automatically trigger data exfiltration to external domains.
[COMMAND_EXECUTION]: The skill uses a set of Task management tools (TaskCreate, TaskUpdate, TaskList) to track progress. It also uses standard filesystem tools (Read, Grep, Glob) for codebase exploration. These tools are used as intended for project planning and do not involve the execution of arbitrary shell commands or privilege escalation attempts.

project-planner