The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The multi-agent workflow involves reading and synthesizing position papers and commentaries created in earlier phases. This architecture creates a vulnerability to indirect prompt injection if any agent-generated content contains instructions designed to manipulate the orchestrator or subsequent agents.
Ingestion points: Position papers from phase-1-divergence/ and commentaries from phase-3-commentary/ are read during the Synthesis (Phase 2), Consolidation (Phase 4), and Final Merge (Phase 6) phases.
Boundary markers: The prompt templates for the Synthesizer, Lead Architect, and Polymath Editor do not include explicit instructions to ignore or isolate potential commands or directives that might be embedded within the user-provided or agent-generated position papers.
Capability inventory: The skill uses Write, Edit, Bash, and Task tools to manage the project files and execution flow, providing a significant capability surface if an injection were to occur.
Sanitization: There are no documented mechanisms for sanitizing or validating the content of the generated papers before they are interpolated into the prompts for subsequent agents.

recursive-synthesis