Skill UI

Get a multi-round adversarial critical review of your research ideas, papers, or experimental results from an external LLM reviewer backend (Codex or manual). Compiles a briefing, runs ultra-reasoning review rounds, and returns a verdict with logical gaps, missing experiments, and narrative weaknesses.

The Santa Method is a multi-agent adversarial verification framework designed to improve output quality by requiring two independent reviewers (Agent B and Agent C) to pass the content before publication. It eliminates single-agent bias and systematically improves outputs through iterative refinement loops until convergence or manual escalation.

This advanced code review skill forces genuine perspective shifts by employing three hostile reviewer personas: the Saboteur (production breaks), the New Hire (maintainability), and the Security Auditor (OWASP-informed). It ensures that all code reviews are critically rigorous, eliminating superficial "Looks Good To Me" approvals and catching blind spots missed by the original author or general AI models. Use it before merging a PR for maximum confidence.

AgentFlow transforms your existing Kanban board (Asana, GitHub, Linear) into a self-managing, autonomous AI development pipeline. It orchestrates multiple AI agents through defined stages (Backlog to Done), enforcing deterministic quality gates, performing adversarial reviews, and tracking costs—all visible within your existing project management UI. It provides crash-proof execution and full observability for complex software development lifecycles.

Facilitates evidence-aware scientific ideation with independent generation, structured discussion, explicit assumptions, transparent evaluation, adversarial review, and decision logs. Use for early-stage research brainstorming and prioritizing candidate directions.

This skill performs a rigorous, adversarial review of implemented code. It verifies that the task implementation is complete, structurally sound, respects defined boundaries, and aligns perfectly with approved requirements and design specifications. Use it after an implementer finishes work, or before final acceptance, to ensure code quality and architectural integrity.

Rigorous verification workflow for LaTeX mathematical proofs. Identifies logical gaps via cross-model adversarial review, fixes each gap with full derivations, and re-reviews until convergence. Generates an audit report. Use for checking proofs and verifying mathematical theory papers.

Perform a two-thread adversarial review: a fresh reviewer writes a strong 200-word rejection memo, then another fresh reviewer defends the paper point-by-point and classifies unresolved issues. Use it before theory paper submission or rebuttal preparation to surface the most damaging possible reviewer attack.

Discipline that subjects every non-trivial decision to a fresh-context adversarial review before it stands. Use when assumptions need cross-examination, plans need stress-testing for hidden failure modes, or correctness matters more than speed. Suited for unfamiliar code, security-sensitive logic, and high-stakes migrations.

An adversarial Quality Assurance (QA) pass designed to deep-dive into any codebase for logic errors, security flaws, and functional bugs. Instead of confirming that code 'looks right,' this skill actively attempts to break the code by simulating hostile inputs, tracing critical data paths, and checking for subtle, real-world correctness issues. Essential for high-stakes, critical code reviews.

This advanced skill performs code review by simulating the documented engineering philosophies of historical figures (e.g., Linus Torvalds, Steve Jobs, Ken Thompson). It moves beyond generic AI findings by grounding critiques in specific, citable principles—such as 'good taste' or 'customer experience.' Ideal for high-stakes PRs, architectural reviews, and ensuring deep, pre-submission code hardening.

The Santa Method is a multi-agent adversarial verification framework designed to improve output quality by requiring two independent reviewers (Agent B and Agent C) to pass the content before publication. It eliminates single-agent bias and systematically improves outputs through iterative refinement loops until convergence or manual escalation.