Audits AI-driven decisions for judgment quality, compliance bias, and manipulation vulnerability.
npx skills add https://github.com/dundas/uhr --skill decision-review-dundas
Please help me install this Agent Skill. Skill: decision-review Source: https://github.com/dundas/uhr/tree/main/.claude/skills/decision-review Command: npx skills add https://github.com/dundas/uhr --skill decision-review-dundas
This Skill addresses the critical need to audit AI-driven decisions, ensuring they are not only technically sound but also free from judgment failures, bias, and manipulation vulnerabilities, thereby closing the "judgment gap" in AI self-improvement loops.
Run a decision review for the current session.
High-intent search queries and answers about installing and using this skill.
To audit AI decisions for bias and compliance, you systematically review them against premise, beneficiary, reversal, skepticism, and alternative checks. This process identifies judgment failures and flags potential compliance red flags in AI sessions.
AI decision auditing is the systematic review of significant AI session decisions to ensure technical soundness and freedom from judgment failures. It is needed after an AI agent deploys code or communicates externally to verify correctness.
Detect prompt injection vulnerabilities by auditing AI decisions for manipulation susceptibility. The review evaluates decisions by type—Action, Architecture, Compliance, Refusal—and flags instances where compliance was prioritized over correctness due to training.
Checking AI agent decisions for helpfulness bias involves running a skepticism check during the decision review. This flags potential issues related to compliance without verification and susceptibility to manipulation.
Decision auditing works for AI code deployment sessions by reviewing the decision-making process post-deployment. It generates findings for Orange and Red decisions and promotes recurring patterns into pre-flight checks and memory logs.
Limitations of automated AI safety checks include the potential for helpfulness bias and compliance without verification. The auditing process mitigates this by applying five key questions to flag red flags related to judgment and scope.
bostonaholicEvaluate proposed architectural decisions and generate Architecture Decision Records.
haabeChallenge core assumptions using pre-mortem, assumption reversal, and red team critiques.
UseAI-proDetect and neutralize prompt injection attacks in text sources.
htafollaAudit and validate framework compliance with detailed audit reports.
