Skill UI

Phoenix Evals enables developers to build and run evaluators for AI and LLM applications. It supports both code-based and LLM-based evaluation methods, including RAG systems. Users can create datasets, run experiments, validate evaluator accuracy, and implement production guardrails. Available in Python and TypeScript, it helps ensure model performance and reliability through systematic testing and human validation.

A persistent, graph-based knowledge base designed to accumulate structured knowledge throughout the entire research lifecycle. It stores and connects diverse entities—including published papers, novel ideas, conducted experiments, and testable claims. Inspired by LLM Wiki patterns, it ensures knowledge compounding, meaning the wiki becomes progressively smarter with every piece of input, rather than requiring re-derivation for every query.

Use the Adaptyv Bio Foundry API and Python SDK to design, submit, and retrieve protein experiments, including binding, thermostability, and expression assays via programmatic workflows.

DeepTools is a comprehensive Python command-line toolkit for processing and analyzing high-throughput sequencing data. It supports quality control, normalization, sample comparison, and generating publication-quality visualizations for ChIP-seq, RNA-seq, and ATAC-seq experiments. Key features include converting BAM to bigWig, assessing sequencing depth, creating heatmaps around genomic features, and performing correlation analysis. Ideal for bioinformatics workflows requiring robust genomic data processing and visualization.

SSH job queue for multi-seed/multi-config ML experiments with OOM-aware retry, stale-screen cleanup, and wave-transition race prevention. Use for batch experiments, grid runs, multi-seed sweeps, or auto-chain experiments when /run-experiment cannot handle 10+ jobs that need orchestration.

Audit experiment integrity before claiming results. Uses an independent cross-model reviewer to detect fake ground truth, score normalization fraud, phantom results, and insufficient experimental scope. Invoke after experiments complete or whenever experimental claims need honest verification.

This tool functions as a post-task epilogue, designed to audit and record the entire research journey. It scans the conversation history at the end of a coding or research session to extract critical research artifacts, including decisions, experiments, dead ends, claims, heuristics, and major pivots. It systematically writes this provenance data into the structured `ara/` directory, ensuring a faithful and auditable trace of how the project truly evolved.

Orchestrates a text-only resubmission of a polished academic paper to a different venue under hard constraints: no new experiments, no bibliography edits, no framework changes, and no overwriting of prior submission directories. It composes the paper into a sibling directory while absorbing prior reviewer concerns through constrained microedits.

This skill guides the safe execution of chaos engineering experiments to surface weaknesses in production systems without causing actual outages. It provides tools to design robust experiments by defining hypotheses, calculating the potential 'blast radius' before failure injection, setting strict abort criteria, and generating structured postmortems. Ideal for SRE teams running game days and resilience tests.

This guide provides a comprehensive framework for designing statistically rigorous A/B tests, split tests, and multivariate experiments. It covers everything from formulating strong hypotheses and calculating sample sizes to selecting appropriate primary, secondary, and guardrail metrics for various business contexts.

This expert framework guides users through the systematic process of conducting rigorous A/B tests and growth experiments. It covers all stages: formulating strong hypotheses, selecting appropriate metrics (primary, secondary, guardrail), calculating necessary sample sizes, and choosing the best implementation method (client-side vs. server-side). Use this when planning to compare two versions or build a systematic growth experimentation program.

Plan experiments before data collection: choose designs, randomize, block, and generate reproducible DOE layouts to avoid confounding and pseudoreplication.