The Agent Skills Directory

[PROMPT_INJECTION]: The skill implements a 'Handoff Convention' (documented in references/harness.md and commands/go.md) which creates a potential surface for indirect prompt injection. The agent is instructed to read a HANDOFF.md file and 'Execute the next action as your first work,' which can include shell commands.\n
Ingestion points: The agent reads from HANDOFF.md (project root) and other documentation files like GOAL.md and DESIGN.md.\n
Boundary markers: While the skill uses delimited markers () for its own modifications to project instruction files, it does not define boundaries or isolation protocols for the content of the HANDOFF.md file itself.\n
Capability inventory: The agent framework expects the capability to execute shell commands, update system instruction files (e.g., CLAUDE.md), and perform broad file system writes based on instructions in the handoff.\n
Sanitization: The instructions lack any requirement for sanitization, validation, or safety checks of the 'next action' commands or instructions provided within the handoff artifacts.\n- [COMMAND_EXECUTION]: The instruction set in commands/go.md explicitly directs the agent to 'Execute the next action' described in HANDOFF.md as its first priority. These actions are defined to include 'exact commands,' facilitating the execution of arbitrary shell commands from potentially untrusted project files without intermediate human verification steps.\n- [SAFE]: The skill performs routine maintenance of project documentation (goals, design decisions, and blueprints) and uses standard delimited blocks to update agent instruction files. These behaviors are properly scoped within the vendor's intended functionality for project management and continuity.

attention-driven