The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill possesses an attack surface for indirect prompt injection by ingesting untrusted data from external sources via web_search and web_fetch (Phase 1d) and from the user during the 'Elenctic Interview' (Phase 1c). This content is directly used to ground the 'Electric Monk' subagents and inform the final synthesis.
Ingestion points: Phase 1d (Parallel research agents), Phase 1c (Socratic interview).
Boundary markers: While the skill uses context files (e.g., context_briefing.md), it lacks specific instructions for sanitizing or escaping potentially malicious instructions embedded in the fetched external content.
Capability inventory: The orchestrator uses web_search, web_fetch, read_file, and spawns sub-sessions with independent tool access.
Sanitization: No explicit sanitization or filtering of external content is described before it is interpolated into monk prompts.
[COMMAND_EXECUTION]: The skill provides templates for spawning subagents using shell-based pipes (e.g., echo "[MONK A PROMPT]" | claude -p). This pattern involves the dynamic assembly of shell commands from generated strings, which presents a risk of command injection if the prompt generation logic includes shell-sensitive metacharacters.
[PROMPT_INJECTION]: The skill instructs the agent to generate subagent prompts that explicitly override standard AI behaviors, such as being balanced or acknowledging opposing views. Instructions like 'Do NOT be balanced. Do NOT acknowledge the other side's merits. BELIEVE.' are intended to ensure the subagents maintain the 'Electric Monk' persona, but they represent a deliberate attempt to bypass the model's default safety and neutrality guidelines for those sub-sessions.

dialectic