DevOps & Infra

ndpvt-web Skill

Decouple evidence access from evidence use when evaluating or building long-context and RAG systems under semantic interference. Use this skill when the user says: 'evaluate my RAG pipeline against hard negatives', 'stress-test retrieval with semantic distractors', 'build a decoupled retrieval benchmark', 'diagnose why my long-context QA is failing', 'create collision-tested evaluation data', 'measure evidence access vs answer quality separately'.

ndpvt-web Skill

Iterative Dual-Phase Financial-PoT: decouple semantic reasoning from arithmetic computation to eliminate calculation errors in financial analysis. Use when: 'calculate financial ratios from reports', 'analyze annual report numbers', 'compute ROE/ROA from statements', 'extract and calculate metrics from financial data', 'why is my financial calculation wrong', 'build a financial analysis pipeline'.

ndpvt-web Skill

Apply the DevOps-Gym methodology to systematically tackle full-cycle DevOps tasks: build/configuration repair, runtime monitoring and anomaly detection, issue resolving via code patches, and regression test generation for Java and Go projects. Trigger phrases: 'fix this build failure', 'diagnose this runtime anomaly', 'generate regression tests for this bug', 'resolve this issue in Java/Go', 'debug this CI pipeline', 'monitor this running service for anomalies'.

ndpvt-web Skill

Build reliable Docker environments for arbitrary code repositories using an agentic, multi-phase approach with dependency reasoning, loop detection, and cross-task success memory. Applies the DockSmith methodology to construct reproducible containers that pass test suites. Trigger phrases: "dockerize this repo", "build a Docker environment for this project", "create a reliable Dockerfile", "set up a containerized dev environment", "fix my Docker build failures", "make this repo's tests run in Docker"

ndpvt-web Skill

Apply Agent Reasoning Reward Model (Agent-RRM) structured critique to improve multi-step agent trajectories. Evaluates tool-use chains with explicit reasoning traces, focused critiques, and process scores. Use this skill when: - "Critique this agent's reasoning trace" - "Evaluate my tool-calling workflow and find flaws" - "Score this multi-step agent trajectory" - "Help me build a reward model for agent training" - "Improve this agent's reasoning with structured feedback" - "Debug why my agent pipeline produces wrong answers"

ndpvt-web Skill

Implement online self-evolving memory for LLM agents using dual-bank architecture (Experience Bank + Meta-Guideline Bank) with reinforcement-weighted retrieval. Use when asked to: 'build an agent that learns from past mistakes', 'add evolving memory to my agent', 'implement experience-weighted retrieval', 'make my agent improve over time from feedback', 'create a self-improving agent pipeline', 'add reinforcement-based memory to my LLM system'.

ndpvt-web Skill

Build memory-augmented time series forecasting systems using hierarchical experience storage (historical patterns, reasoning wisdom, general laws) with LLM-driven inference and reflective iteration. Trigger phrases: 'forecast time series with memory', 'experience-driven prediction', 'memcast forecasting', 'hierarchical memory for time series', 'LLM time series with retrieval', 'memory-augmented forecasting pipeline'

ndpvt-web Skill

Defend against and red-team physical prompt injection attacks on Large Vision-Language Models (LVLMs). Build input sanitization pipelines, attention-aware OCR filters, and robustness test harnesses that detect typographic instructions embedded in images before they reach the model. Trigger phrases: "harden my vision model against prompt injection", "detect text injection in images", "red team my LVLM pipeline", "build an image input sanitizer for VLMs", "test physical prompt attacks", "PPIA defense pipeline"

ndpvt-web Skill

Build self-improving agent systems that distill raw execution traces into a hierarchical skill library (SkillBank) and recursively evolve those skills based on failure analysis. Use when asked to: 'build an agent that learns from mistakes', 'create a skill library from past runs', 'improve agent performance over iterations', 'extract reusable patterns from logs', 'reduce prompt bloat from trajectory history', 'make my agent pipeline self-improving'.

ndpvt-web Skill

Build structured narrative memory systems from conversational traces using TraceMem's three-stage pipeline (segmentation, consolidation, clustering). Use when asked to: 'build a memory system for a chatbot', 'implement long-term conversation memory', 'create user memory cards from chat history', 'add persistent memory to an LLM agent', 'organize conversation history into structured narratives', 'implement episodic memory with clustering'.

ndpvt-web Skill

Build and deploy real-time open-vocabulary instance segmentation pipelines using YOLOE-26, which combines YOLOv26's NMS-free architecture with YOLOE's open-vocabulary embedding heads. Covers text-prompted, visual-prompted, and prompt-free segmentation via the Ultralytics API. Trigger phrases: - "Set up YOLOE-26 for open-vocabulary segmentation" - "Detect and segment arbitrary object classes in real time" - "Build a visual prompt segmentation pipeline with YOLO" - "Deploy open-vocabulary instance segmentation with TensorRT" - "Segment objects by text description without retraining" - "Use YOLOE for prompt-free object detection"

ndpvt-web Skill

Build generative NER systems using LLMs with optimal output formats and prompt engineering. Use when: 'extract entities from text', 'build a NER pipeline with an LLM', 'named entity recognition with generative models', 'format NER output as XML or bracketed', 'fine-tune a model for entity extraction', 'nested entity recognition'.

ndpvt-web Skill

Automatically discover and test for hidden biases in LLM-as-a-Judge evaluation pipelines using the BiasScope framework. Generates bias hypotheses, perturbs test cases, and validates whether judge models are susceptible. Use when: 'audit my LLM judge for bias', 'find biases in my evaluation pipeline', 'stress test my LLM evaluator', 'check if my model judge is robust', 'discover unknown biases in my scoring system', 'build adversarial eval sets for my judge'.

ndpvt-web Skill

Build cost-efficient RAG pipelines for entity matching and deduplication using blocking-based batch retrieval and generation. Reduces LLM API calls and latency by grouping similar entity pairs into blocks before retrieval and inference. Use when the user asks to 'match entities across datasets', 'deduplicate records with LLMs', 'build a RAG pipeline for entity resolution', 'reduce cost of LLM-based record matching', 'link records between two tables', or 'entity matching with knowledge augmentation'.

ndpvt-web Skill

Build multi-agent systems for end-to-end ontology curation from scientific literature. Applies FlyAOC's agent architecture patterns—memorization, pipeline, single-agent, and multi-agent—to extract structured, ontology-grounded annotations from document corpora. Use when asked to: 'curate knowledge from papers into a structured ontology', 'build an agent pipeline for scientific literature extraction', 'design a multi-agent system for document annotation', 'extract Gene Ontology or controlled-vocabulary terms from text', 'reconcile evidence across multiple documents into structured annotations', 'build a retrieval-augmented scientific reasoning system'.

ndpvt-web Skill

Build and evaluate multilingual machine reading comprehension systems for HR documents (resumes and job descriptions). Implements the JobResQA pipeline: tiered QA generation, cross-document reasoning, placeholder-based bias testing, TEaR translation, and G-Eval LLM-as-judge scoring. Use when: 'evaluate resume parsing accuracy', 'build HR question answering', 'test multilingual resume understanding', 'check bias in resume screening', 'cross-document QA on resumes and JDs', 'benchmark LLM on HR tasks'.

ndpvt-web Skill

LLM-assisted translation annotation: build parallel corpus annotation pipelines with template-based prompt management, structured JSON output constraints, stand-off annotation architecture, and human-in-the-loop review for sentence alignment, word alignment, translation shift classification, and error taxonomy. Trigger phrases: "annotate this translation", "align these parallel sentences", "classify translation shifts", "build a translation annotation pipeline", "create parallel corpus annotations", "translation quality annotation"

ndpvt-web Skill

Generate accurate SQL from natural language using the SSEV pipeline: schema-linked prompting, execution-guided self-refinement, and weighted majority voting across multiple candidate queries. Use when the user says 'write SQL for this question', 'query this database', 'convert this question to SQL', 'text-to-SQL', 'generate a SQL query from natural language', or 'help me query this schema'.

ndpvt-web Skill

Defensive dual-use risk assessment for open science artifacts. Evaluates research papers, datasets, methods, and tools for repurposing vulnerabilities using a structured pipeline based on Hashemi et al. (2026). Produces risk reports with harmfulness, feasibility-of-misuse, and technical-soundness scores. Trigger phrases: "assess dual-use risk of this paper", "evaluate artifact repurposing risk", "run dual-use risk audit on this dataset", "check this tool for misuse potential", "security review of open science artifacts", "red-team this research for repurposing vulnerabilities"

ndpvt-web Skill

Build knowledge-graph-aided ASR error correction pipelines for medical speech, using phonetic similarity + semantic retrieval to fix misrecognized terminology before LLM reasoning. Use when: 'correct medical ASR errors', 'fix speech recognition for clinical terms', 'build medical speech QA pipeline', 'knowledge graph for ASR correction', 'phonetic matching for medical terms', 'improve Whisper transcription of medical audio'.

ndpvt-web Skill

Build autonomous multi-scale scientific simulation pipelines using the QUASAR architecture: a Strategist-Operator-Evaluator agent trio with adaptive planning, hierarchical knowledge retrieval, and context-efficient memory. Triggers: 'build an autonomous simulation pipeline', 'orchestrate DFT and MD workflows', 'QUASAR-style agent system', 'multi-scale atomistic workflow', 'autonomous computational chemistry', 'scientific simulation agent framework'

ndpvt-web Skill

Detect AI-generated images using NLM noise entropy analysis and build robust forensic detection pipelines. Use when: 'detect if an image is AI-generated', 'build an AI image detector', 'classify real vs fake images', 'image forensics pipeline', 'noise entropy feature extraction', 'detect deepfakes or inpainted regions'.

ndpvt-web Skill

Implement pipeline-aware intermediate representation (IR) caching for agentic systems. Instead of caching final LLM responses, decompose multi-step pipelines into stages and cache structured reasoning artifacts at each checkpoint. Triggers: 'cache intermediate reasoning', 'reduce redundant LLM calls', 'pipeline-aware caching', 'agentic caching strategy', 'cache reasoning not responses', 'structured IR caching for agents'

ndpvt-web Skill

Audit Ethereum smart contracts for business logic vulnerabilities using Heimdallr's four-phase agentic pipeline: function-level code reorganization via dependency graph clustering, heuristic Plan-Remind-Solve reasoning with adversarial state injection, automatic multi-step exploit chaining, and cascaded false-positive filtration. Trigger phrases: 'audit this smart contract', 'find vulnerabilities in this Solidity code', 'check this DeFi protocol for exploits', 'smart contract security review', 'detect business logic bugs in this contract', 'chain exploit paths in this protocol'.

ndpvt-web Skill

Build query-adaptive knowledge graph retrieval systems using CatRAG's context-aware traversal. Transforms static KG-based RAG pipelines into dynamic, query-sensitive retrieval that recovers complete multi-hop evidence chains. Use when: 'build a multi-hop RAG pipeline', 'improve knowledge graph retrieval', 'fix semantic drift in graph search', 'implement context-aware graph traversal', 'retrieve complete evidence chains from a KG', 'add query-dependent edge weighting to my graph'.

pku-yuangroup Bundle

Aggregates per-tool QC metrics (FastQC, fastp, alignment, quantification, variant calling, single-cell) into one interactive MultiQC report, and guides module scoping, sample-name resolution, large-cohort behavior, and turning the report into an actual QC gate. Use when summarizing QC across many samples, building a shareable quality report, or wiring automated QC into a pipeline.

pku-yuangroup Bundle

End-to-end shotgun metagenomics workflow from FASTQ to taxonomic and functional profiles, orchestrating controls/host depletion, Kraken2+Bracken classification, MetaPhlAn marker profiling, and HUMAnN functional profiling. Covers the controls-first ordering, why Kraken2 read counts are not abundances and MetaPhlAn cell fractions do not equal Bracken read fractions, and the consistent-pipeline framing. Use when profiling shotgun metagenomic samples end to end, or chaining classification, abundance, and function. For resistome see metagenomics/amr-detection; for strains see metagenomics/strain-tracking; for assembly see genome-assembly/metagenome-assembly.

pku-yuangroup Bundle

End-to-end pooled and single-cell CRISPR screen analysis from FASTQ to hit genes. Orchestrates library design QC, guide counting, six-stage screen QC (plasmid Gini, replicate Pearson, CEGv2 PR-AUC, copy-number artifact), method-appropriate hit calling across MAGeCK RRA/MLE, BAGEL2, drugZ, JACKS, and Chronos, cancer-cell-line copy-number correction (CRISPRcleanR / Chronos), batch correction for multi-batch screens, and the specialized branches for combinatorial paralog screens, single-cell Perturb-seq, base-editor variant-function screens, prime-editor screens, and in vivo bottleneck-aware screens. Use when analyzing any pooled CRISPR screen end-to-end, matching the hit-calling method to the experimental design, integrating copy-number correction into the pipeline, or branching the workflow for single-cell, combinatorial, base-editor, prime-editor, or in vivo variants.

pku-yuangroup Bundle

Download genome assemblies, gene records, and ortholog data from NCBI using the modern Datasets v2 CLI (replaces assembly_summary.txt scraping and many EFetch workflows). Use when bulk-pulling genome assemblies, gene metadata across species, ortholog sets, or BLAST databases; when E-utilities are too slow for genome-scale work; or when automatic checksum verification, parallel download, and clean accession-driven retrieval are required. Encodes the JSON-lines output format, dataformat conversion, --dehydrated for cloud workflows, and when Datasets is/isn't the right tool.

pku-yuangroup Bundle

Stores and operates on sparse expression matrices for single-cell and large bulk RNA-seq, covering dgCMatrix/dgRMatrix/dgTMatrix when-each-is-fast, the dgCMatrix (CSC, R) <-> CSR (Python) implicit transpose, AnnData (cells-rows) <-> SingleCellExperiment (cells-cols) orientation flip, HDF5/h5ad vs Zarr cloud-native shift, HDF5SummarizedExperiment + DelayedArray for out-of-memory bulk, scanpy backed mode for large h5ad, the ~10-15% density crossover where dense beats sparse, 10X format proliferation (MTX vs CellRanger H5 vs h5ad), the dense-conversion memory blow-up, and Dask + Zarr for consortium-scale matrices. Use when choosing sparse format, working with single-cell-sized matrices, importing/exporting 10X, debugging R/Python interop transposes, processing matrices too large for RAM, or building cloud-native pipelines.

pku-yuangroup Bundle

Authors bioinformatics pipelines in WDL (Workflow Description Language) run by Cromwell or miniwdl, targeting the GATK/Broad and Terra/AnVIL/BioData Catalyst cloud ecosystem, with tasks, workflows, scatter-gather parallelism, structs, and a runtime block that sizes the cloud VM. Use when deciding to target Terra/AnVIL/GATK/WARP (chosen for the ecosystem, not the language); sizing runtime disks dynamically for a fresh-per-task cloud VM (ceil(size(f)*factor)+buffer); choosing preemptible vs on-demand VMs by task length and idempotency; picking Cromwell (production, cloud, call-caching) vs miniwdl (local dev, miniwdl check linting, readable errors); enabling and debugging call-caching silent-miss modes; pinning Docker by digest for reproducibility and cache stability; or scattering an array for parallel fan-out.

pku-yuangroup Bundle

End-to-end clinical trial analysis workflow from CDISC SDTM/ADaM loading through ICH E9(R1) estimand-driven primary analysis to CONSORT 2025 regulatory-compliant reporting. Covers data preparation, FDA 2023 marginal vs conditional logistic regression, categorical tests with Boschloo, modern HTE/subgroup methods, missing-data sensitivity (MMRM, reference-based MI, Permutt tipping point), graphical multiplicity (Bretz-Maurer), survival analysis (Cox/RMST/competing risks) when applicable, and Table 1. Use when performing a complete analysis of clinical trial data.

pku-yuangroup Bundle

Orchestrates an end-to-end CRISPR editing experiment design from target gene to delivery-ready, validatable constructs. Sequences guide design, off-target assessment, edit-modality selection (knockout, base editing, prime editing, HDR knock-in), and template/donor design, with a QC checkpoint at each handoff. Use when designing a complete CRISPR experiment for knockout, point correction, or tagging and the order of operations, the modality decision, and the cross-cutting traps are needed rather than a single step. Defers each step's mechanics to the genome-engineering skills.

pku-yuangroup Bundle

Detects m6A modifications from Oxford Nanopore direct-RNA-seq (DRS) signal using m6Anet (multiple-instance-learning over DRACH 5-mer signal). Covers the upstream pipeline (Dorado/Guppy basecalling -> minimap2 map-ont -> nanopolish eventalign -> m6anet dataprep -> m6anet inference), per-site vs per-read probability including the mod_ratio stoichiometry column, the DRACH-only constraint, minimum-coverage thresholds (20-50 reads/site), multi-condition comparison via xPore/Nanocompore/ELIGOS, Dorado native modification calling (RNA004, 2024+), and the cDNA-vs-DRS distinction (cDNA Nanopore CANNOT detect modifications). Use when calling m6A from ONT DRS without immunoprecipitation, choosing m6Anet vs xPore vs Nanocompore vs ELIGOS vs Dorado native, interpreting probability_modified vs mod_ratio vs per-read probabilities, deciding between m6Anet (known DRACH sites) and Dorado/Remora (genome-wide screening), pinning RNA002 vs RNA004 chemistry and basecaller versions, or troubleshooting eventalign/dataprep failures.

pku-yuangroup Bundle

End-to-end post-GWAS causal inference pipeline orchestrating heritability partitioning, genetic correlation, Mendelian randomization with CHP-aware sensitivity (CAUSE / LHC-MR), colocalization, fine-mapping with SuSiE / FOCUS, mediation, TWAS triangulation, cis-pQTL drug-target MR, effector-gene prioritization (L2G / PoPS / cS2G), and GenomicSEM common-factor GWAS. Use when triangulating causal inference across multiple complementary methods, prioritizing tissues via stratified LDSC, nominating or de-risking drug targets, mapping a lead SNP to a candidate effector gene, modeling shared genetic architecture across correlated traits, or producing a STROBE-MR-compliant publication-grade evidence battery from GWAS summary statistics.

pku-yuangroup Bundle

Orchestrates an end-to-end de novo genome assembly project, routing each step to the right genome-assembly skill rather than restating it. Profiles the genome first (k-mer spectrum -> size, heterozygosity, ploidy), QCs reads, chooses an assembly path by data type (SPAdes for Illumina, Flye for noisy long reads, hifiasm for HiFi, metaFlye for communities), polishes only when needed, decontaminates, scaffolds with Hi-C, and finishes with three-axis QC (contiguity + completeness + correctness). Use when assembling a genome from raw reads and deciding which assembler, whether to polish, and how to prove the result is good.