The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses platform-specific cron tools (CronCreate) to establish a persistent, recurring autonomous loop that executes every 10 minutes independently of the active session.
[PROMPT_INJECTION]: The skill processes untrusted project files to drive its autonomous execution loop, creating a surface for indirect prompt injection.
Ingestion points: Reads CLAUDE.md, README.md, and project source files during the ORIENT and Phase 0 phases to determine goals and next steps.
Boundary markers: While the skill uses an 'Execution Protocol' to maintain its internal rules, it lacks explicit delimiters or instructions to treat external file content as untrusted data during subagent reasoning.
Capability inventory: Possesses significant capabilities including arbitrary command execution in the ACT phase, file writes to state and notebook files, and the management of its own cron-based persistence.
Sanitization: No evidence is provided for the sanitization or validation of instructions ingested from the analyzed project documentation.
[PROMPT_INJECTION]: Implements a self-instruction mechanism termed an 'Execution Protocol' or 'bootstrap sector' within PROJECT_STATE.md, designed to override or reconstruct the skill's operational rules if the agent's internal context is lost.
[COMMAND_EXECUTION]: Executes dynamically generated shell commands during the ACT phase based on decisions made in previous phases, including running dev servers and build tools.
[COMMAND_EXECUTION]: Modifies the agent's platform environment by creating a hidden configuration file (~/.claude/.memento-auto) without a specific user-facing prompt or toggle.

liquid-cat-physics