PromptAutopsy

PromptAutopsy is a web app (with optional desktop agent) that helps teams debug LLM behavior by turning messy prompt iterations into structured, testable hypotheses. You paste a prompt, model outputs, and any relevant context; the app automatically generates a minimal test suite (counterfactual prompts, paraphrases, constraint toggles) and runs them across selected models. It then produces an “interpretability-lite” report: which instruction segments correlate with failures, where the model ignores constraints, and what context tokens likely caused derailment. It’s not trying to be academic mechanistic interpretability; it’s practical root-cause analysis for product teams shipping LLM features. Reports are exportable for audits and incident reviews, and the app tracks regressions over time as prompts, tools, and models change. Expect some false positives—this is decision support, not ground truth.