The Agent Skills Directory

[PROMPT_INJECTION]: The skill instructions enforce strict output formatting ('No preamble. No postamble. No explanations. No filler/hedging/transitions.'). While framed as a token-saving measure, this constraint is a common technique used to suppress an agent's built-in safety explanations and guardrails, potentially allowing for the generation of harmful content without user warning.
[PROMPT_INJECTION]: Eight files in the references/ directory (including architecture-patterns.md, code-organization.md, and others) contain 150 sections each of nearly identical, repetitive technical filler text. This 'context stuffing' creates an adversarial environment that can bury malicious instructions in a massive volume of data, or be used to exhaust the agent's attention and context window limits.

data-science-causal-inference