The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user-authored content, such as task definitions and prompts, which creates a surface for indirect prompt injection.
Ingestion points: The skill accepts task_source and task_id inputs in SKILL.md and references/seed-fidelity-checklist.md to analyze user tasks.
Boundary markers: The workflow relies on standard markdown headers (e.g., ## Prompt, ## Success Criteria) to distinguish data types but does not implement hardened delimiters against adversarial content.
Capability inventory: The skill utilizes tools for task state verification (verify_seed), deterministic evaluation (evaluate_criteria), and sandbox creation (run_task), and it documents the use of shell commands (curl) for probing twin environments.
Sanitization: The instructions mitigate risk by requiring sensitive session tokens (agent_token) to be handled via environment variables and by restricting diagnostic network operations to read-only (GET) requests.
[SAFE]: The skill demonstrates strong security awareness regarding credential management. It explicitly instructs the agent and user not to paste live credentials into command history, messages, or files, and provides a clear workflow for discarding sandbox sessions via a mandatory discard-token mechanism.

pome-verify-seed