Skill UI

This comprehensive pipeline guides users from a vague research concept to a fully structured, testable academic proposal. It first refines the core methodology (research-refine) to establish a stable thesis, and then generates a detailed, claim-driven experiment roadmap (experiment-plan). This end-to-end workflow ensures that the final academic output is both elegant and empirically validated.

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.

Provides patterns and techniques for building robust evaluation loops for AI agents. This skill guides the implementation of self-critique, evaluator-optimizer pipelines, and test-driven refinement workflows. Use it to significantly enhance the accuracy and quality of complex outputs like code, reports, and analyses by implementing structured, iterative improvement cycles (Generate → Evaluate → Critique → Refine).

This skill provides a comprehensive framework for optimizing, improving, and debugging Large Language Model (LLM) prompts. It leverages production trace data, system evaluations, and human annotations to extract prompts from spans. Users can run a data-driven optimization loop, ensuring the AI responses meet desired quality standards. Ideal for prompt tuning, system prompt refinement, and improving overall AI output quality.

Interactive workflow that interrogates scope, deliverables, constraints, and success criteria while guiding users to refine task prompts; copies the final markdown with Joyride clipboard commands and never writes code.

Interactive workflow that uses the joyride_request_human_input tool to probe scope, deliverables, and constraints before executing any task, ensuring thorough clarification and structured planning.

Generate professional infographics with Nano Banana Pro AI, Gemini quality review, and optional Perplexity research. Supports 10 infographic types, 8 industry styles, colorblind-safe palettes, and smart iterative refinement.

This skill facilitates a rigorous, documentation-anchored design review session. It challenges a proposed plan by cross-referencing it against the project's established domain glossary (CONTEXT.md), historical architectural decisions (ADRs), and the current codebase. It forces the user to resolve ambiguities, define precise terminology, and validate design trade-offs using concrete scenarios, ensuring all decisions are captured and documented inline.

This skill autonomously optimizes artifacts like code or models using Hypothesis Tree Refinement. It runs iterative experiments with dev/test evaluators to prevent overfitting, coordinating subagents in isolated worktrees. Ideal for improving training recipes, agents, or pipelines without manual supervision.

This skill acts as a central router, helping users determine the correct specialized skill or workflow to tackle development challenges. It guides the entire software development lifecycle, from initial idea refinement (Idea -> Ship) and prototype testing, to managing complex multi-session builds and handling incoming bugs via triage. It is essential for navigating the vast ecosystem of AI coding agents.

Use this powerful skill to perform advanced, multi-modal video generation and editing using the official google-genai SDK. Capabilities include generating realistic videos from text prompts (Text-to-Video), animating scenes from a single input image (First-Frame-to-Video), and applying style or character references from images. It also supports complex video refinement, inpainting, and optimizing source footage for professional creative workflows.

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.