The Agent Skills Directory

[PROMPT_INJECTION]: The skill processes external documents (PDF/Word) through the scheme_extractor subagent to define proposal structures and evaluation criteria. This creates a vulnerability to indirect prompt injection, where malicious instructions embedded in a user-provided guideline could override the agent's behavior during the drafting or review phases.
Ingestion points: agents/scheme_extractor.md reads user-provided funding guidelines or sample forms.
Boundary markers: Absent; there are no specific delimiting instructions or warnings to ignore adversarial content within the extracted data.
Capability inventory: agents/compliance_gate.md executes local Python scripts; the pipeline invokes downstream drafting and review skills.
Sanitization: Absent; the skill does not filter or validate the textual content extracted from uploaded documents before using it to guide agent logic.
[COMMAND_EXECUTION]: The skill performs shell command execution to run utility scripts located in the local scripts directory.
Evidence: agents/compliance_gate.md and references/format-and-limits.md instruct the agent to run python3 scripts/proposal_wordcount.py <draft>.md --scheme <id> and scripts/privacy_scan.py to enforce compliance and data privacy.

food-proposal