The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill acts as a central hub (Convener), gathering and processing untrusted data from user tasks, document corpora, and web search results (via WebSearch and WebFetch), which is then interpolated into prompts for downstream agents. Ingestion points include user prompts, external search results, and the project wiki. Boundary markers such as '=== EVIDENCE BRIEF ===', '## Standing instructions', and '## P1 wiki context' are used to separate data types. Capability inventory includes agent spawning, local file reading, network searches, and shell command execution. Sanitization is provided by a 'safeguarding-reviewer' agent and an ethical pre-screen for sensitive data instruments.
[DATA_EXFILTRATION]: The skill contains an 'External Critic' feature (Phase 5.5) that allows sending deliverable text to external AI models for critique. Although this process includes a sensitivity gate and requires explicit user approval, it creates a mechanism for sensitive project data to leave the local environment. Additionally, the skill uses absolute Windows file paths (e.g., 'C:/Users/AGasser/OneDrive/...') to access project files like 'programme_context.json', which could expose organizational data if the agent's context is compromised.
[COMMAND_EXECUTION]: The skill instructions specify the execution of local Python scripts ('scripts/cross_model_critic.py') and specific Python modules ('ane_package.qa.word_count') via the shell. It also utilizes Git commands ('git log', 'git check-ignore') to perform recovery checks and manage file states.

ann