super-review:llm-prompts

Reviewing LLM prompt templates for prompt injection risks involves auditing PR diffs to flag unsafe patterns where data and instructions merge via delimiters or interpolations. This process identifies injection-prone shapes within prompt files to prevent security vulnerabilities.

Structured output schema validation for LLM prompts is the process of enforcing explicit output schemas, length caps, and stable structured shapes in prompt content. It identifies prompts lacking these constraints to prevent parsing failures and ensure machine-readable responses.

Adding few-shot examples and eval datasets to prompt templates requires ensuring custom formats have worked examples and pairing prompt changes with regression datasets. This checks few-shot coverage and verifies prompts have machine-checkable assertions for evaluation hygiene.

Yes, the prompt review checks for PII leakage in few-shot examples by detecting real personally identifiable information within prompt content. It also identifies prompts requesting chain-of-thought reasoning that may be returned verbatim, ensuring reasoning hygiene.

Automated prompt content reviews are limited to matching paths under prompts/ and eval/ directories, plus long system-prompt literals in code. They focus on anti-patterns causing injection risk, parsing failures, ambiguity, non-determinism, and eval regressions within those specific PR diff boundaries.