OmarAlex24·OmarAlex24/audit-skills/skills/llm-app-review/SKILL.md

llm-app-review

Skill1 starsChanged 2 days agoLicence not identifiedView on GitHub

Read-only, line-by-line review of the code of an app that uses an LLM, treating the model as untrusted, expensive and non-deterministic. Looks for outside text reaching the prompt (direct and indirect prompt injection) and what the model does next, tools with excess permissions, model output used unvalidated, uncapped cost (unlogged retries, timeouts that do not cancel, unbounded loops), prompts asking for what the code does not do, model evaluations that are not comparable, secrets or PII sent to the provider or logs, goals met to the letter but not in intent, and provider outages passed off as valid responses. Each finding brings path:line, sequence and a concrete input. Reports and stops. Use it for "review my LLM app", "prompt injection audit", "is my agent safe with tools". Not for RAG design or model choice (architecture-review). Also triggers on Spanish requests such as "revisa el uso del LLM", "¿mi agente es vulnerable a prompt injection?" or "¿cuánto gasta mi app en el modelo?".

• Reads credentials