Machine Learning Skills

Research external context for a dataset — domain background, history, related studies, and why this data matters. Outputs detective.json (structured findings with det_xx IDs) before any analysis begins.

Find a dataset for a Data2Story blog. Accepts a topic, a URL, or a DIP-style category. Downloads + validates against 4 completeness gates before handing off to /data2story-pro. Local-first: searches Economist/Pudding/TidyTuesday clones before going online. Supports --validate-only to audit a folder you already have.

用于审阅人物传记。适合检查身份归属、动机代价、能力边界、关系网、写作抓手、心理弧线与索引映射是否成立，并输出可落库、可回写的传记审阅报告。内置 Nuwa 深度审阅（心智模型、表达DNA、矛盾张力、诚实边界、决策启发式）。关键词：审人物传记、人物审阅、能力边界、关系网、心理弧线、传记报告、心智模型、表达DNA。

用于审阅章节正文的执行质量。适合检查章首抓力、中段回报、章末钩子、现实落地、规则边界、链路失配与可回写的审阅结论。内置人物执行审计（声口一致性、心智模型落地、压力反应匹配、误判/盲区触发、关系拉扯兑现）。关键词：审这章、章节审阅、章首抓力、中段回报、章末钩子、审阅报告、人物执行审计、声口检测。

> Use when the user has a known, already-observed anomaly in their data — a metric spike or drop, an outlier, an unexpected number — and wants its root cause diagnosed, not guessed. Forms a slate of candidate causes, tests each against the data, and eliminates the ones the data refutes, narrowing the live candidates until exactly one survives refutation and passes a positive confirming test. The result is an investigation log with the confirmed root cause and the evidence that ruled out the alternatives. Not for open-ended discovery over a dataset with no specific anomaly in hand (that is data-analysis), and not for checking an external claim against sources (that is claim-verify) — this is reactive diagnosis of one anomaly you already know about.

> Use when the user wants to evolve an ML model/program through population-based search rather than a single sequential refine loop — a generational evolution where parallel proposers each apply one small SEARCH/REPLACE diff to a parent, scored by a cascade-evaluated training run, and children are kept in a MAP-Elites archive across islands (with migration + checkpointing) so diverse high performers survive. A finite, bounded-parallelism re-creation of AlphaEvolve/OpenEvolve, bent for ML autoresearch. Runs to a fixed compute budget or until interrupted. Not for the sequential single-thread autoresearch loops (one change → measure → keep/revert), and not for verifying a known bug or external claim — this is parallel, diversity-preserving search over a program.

> Use when the user has a messy tabular data dump (CSV/TSV/parquet/Excel/JSON) and wants it iteratively cleaned to an inferred data contract — a checklist of deterministic pass/fail checks, not a quality score. A single agent profiles the table, synthesizes a per-column contract compiled into binary checks (types, nulls, duplicates, inconsistent categories, format/range violations, outliers), then applies one targeted transform at a time, keeping it only if it reduces its target check's violations with no regression and no guardrail breach. Stops deterministically when every check passes, every remaining check is an unfixable residual, or a budget is hit; emits a replayable pipeline and an auditable ledger. Not for open-ended analysis of an already-clean dataset, diagnosing one known anomaly, or verifying a claim against sources — those are analytical loops; this rewrites the data to a contract.

Master the art of podcast interviewing using NPR training methodology and Tim Ferriss's preparation techniques to extract compelling stories and insights from any guest. Use when: Preparing for a podcast interview with a guest; Designing questions that elicit stories, not soundbites; Struggling to get guests to open up authentically; Planning a new interview-format podcast; Improving your interviewing technique

Master April Dunford's 8-step sales narrative from \"Sales Pitch\" (2023). Transform your positioning into a compelling story that wins enterprise deals. Use when: Structuring B2B sales presentations and demos; Creating pitch decks for complex products; Training sales teams on narrative-driven selling; Converting positioning strategy into sales conversations; Winning against entrenched competitors

Train machine learning models end-to-end, covering data loading, preprocessing, architecture selection, training loops, validation, and checkpointing.

Retrieves and synthesizes relevant information from a knowledge base using RAG techniques to provide grounded context for a given query.

Clean and preprocess datasets by handling missing values, removing duplicates, correcting types, resolving outliers, and enforcing validation schemas.

Automatically infer loop invariants for code verification and correctness proofs. Use when analyzing loops to identify properties that hold throughout execution, generating assertions for verification, proving loop correctness, or documenting loop behavior. Supports Python, Java, C/C++, and language-agnostic analysis. Generates invariants as code assertions (assert statements). Triggers when users ask to infer invariants, find loop properties, generate loop assertions, prove loop correctness, or verify loop behavior.

Automatically infer formal correctness properties from Verilog/SystemVerilog RTL code and generate SystemVerilog Assertions (SVA). Identifies control-flow invariants (mutual exclusion, valid-ready handshakes, pipeline ordering, safety properties), liveness expectations, and temporal properties. Use when working with RTL designs that need formal property generation, when adding assertions to existing RTL, or when users ask to infer properties, generate assertions, or create formal specifications from hardware designs.

Builds a portable, embedding-free knowledgebase from a set of files and delivers it as a self-contained `.skill` bundle (BM25 index + bundled searcher + query protocol). Use when a user wants to turn uploaded files, a folder, or a corpus into a searchable knowledgebase they can hand to any agent — phrased as "make a knowledgebase", "build a KB skill", "package these docs for retrieval", "create a searchable bundle", or references to a `.skill` KB. The output runs anywhere with Node or Python — no model, no install, no network. Distinct from `bm25` (ephemeral in-session search) and `building-github-index` (markdown project-knowledge index).

Query a portable, embedding-free lexical knowledgebase bundled as a `.skill`. Use when the user references this KB or asks a question whose answer is in its corpus ({{SOURCE}}). Retrieval is BM25 over a precomputed inverted index — there is no embedding model, so YOU expand the query into search terms before searching. Bundle holds index.json + chunks.jsonl + search.js + search.py; pure stdlib, no install, no network.

Zero-shot univariate time series forecasting using the Reverso foundation model (NumPy/Numba CPU-only inference). Activate when users provide time series data and request forecasts, predictions, or extrapolations. Supports Reverso Small (550K params). Triggers on "forecast", "predict", "time series", "Reverso", or when tabular data with a temporal dimension needs future-value estimation.

Execute programs on a compiled transformer stack machine where every instruction fetch and memory read is a parabolic attention head. Demonstrates that transformer attention + FF layers can implement a working computer. Use when user mentions "llm-as-computer", "lac", "stack machine", "compiled transformer", "percepta", "parabolic attention", "execute program", or asks to run/trace programs on the transformer executor.

Analyze AI/ML technical content (papers, articles, blog posts) and extract actionable insights filtered through enterprise AI engineering lens. Use when user provides URL/document for AI/ML content analysis, asks to "review this paper", or mentions technical content in domains like RAG, embeddings, fine-tuning, prompt engineering, LLM deployment.

In-process semantic search over text files or in-memory strings, using Gemini embeddings via the CF AI Gateway. Use when user wants fuzzy/conceptual search where exact-keyword grep would miss — "sessions discussing regulatory constraints", "code about retry logic", "notes mentioning burnout even if the word isn't there". Complements searching-codebases (regex/AST) and extracting-keywords (YAKE). Do NOT use when an exact string/regex match is what's wanted — grep/rg wins on speed and precision there.

ワークスペース内の遠い知識や活動をつないで、新しいアイデアを生成する思考支援スキル。過去記事・メモ・RAG・発信資産から企画や実装ネタを出したい時に使用。「アイデア出して」「知識をつないで」「一見関係ないものをつないで」で使用。

ワークスペース全体をベクトル検索＋構造化ファクト管理する任意スキル。ユーザーがRAGのセットアップ・検索・ファクト操作を明示的に依頼した場合に使う。「RAGをセットアップして」「RAGで探して」「ファクト登録」で使用。

Use when defining or implementing Go interfaces, designing abstractions, creating mockable boundaries for testing, or composing types through embedding. Also use when deciding whether to accept an interface or return a concrete type, or using type assertions or type switches, even if the user doesn't explicitly mention interfaces. Does not cover generics-based polymorphism (see go-generics).

Use when optimizing Go code, investigating slow performance, or writing performance-critical sections. Also use when a user mentions slow Go code, string concatenation in loops, or asks about benchmarking, even if the user doesn't explicitly mention performance patterns. Does not cover concurrent performance patterns (see go-concurrency).

Create a new evaluation dataset or add cases to an existing one for the Azure SDK QA bot evaluation. WHEN: "add eval dataset item", "add a test case", "new evaluation dataset", "create dataset", "add question to dataset", "curate eval data", "promote staging cases", "upload dataset asset", "new scenario dataset". DO NOT USE FOR: running evaluations, pipeline troubleshooting, knowledge-graph indexing.

TRIGGER THIS when researching market salaries, building compensation strategies, creating salary bands, analyzing pay equity, conducting salary reviews, preparing compensation proposals, or ensuring competitive compensation. Analyzes market data, builds equity-focused salary structures, identifies pay gaps, and creates data-driven compensation recommendations.

TRIGGER THIS when building DEI strategy, creating diversity goals, designing inclusive programs, building equity analyses, developing training, addressing inclusion issues, or tracking DEI metrics. Develops comprehensive diversity, equity, and inclusion strategies aligned to business goals with measurable outcomes, training programs, metrics frameworks, and accountability systems.

TRIGGER THIS when preparing interviews, planning hiring processes, creating interview questions, building evaluation frameworks, developing interviewer guides, or standardizing interview approaches. Builds complete interview kits with competency-based questions, scorecards, rubrics, interviewer training, and structured evaluation protocols.

Generate project status reports with RAG status, milestones, risks, blockers, resource utilization, and executive summaries

Profile and explore datasets to understand their shape, quality, and patterns before analysis. Use when encountering a new dataset, assessing data quality, discovering column distributions, identifying nulls and outliers, or deciding which dimensions to analyze.

Design customer onboarding playbooks with milestone tracking, training plans, health checks, and success criteria

> ESM2 protein language model for embeddings and sequence scoring. (2) Getting protein embeddings for clustering, (3) Filtering designs by sequence plausibility, (4) Zero-shot variant effect prediction, (5) Analyzing sequence-function relationships. For structure prediction, use chai or boltz. For QC thresholds, use protein-qc.

Query NCBI GEO for gene expression datasets. Use when user asks about RNA-seq datasets, microarray data, expression data, GEO accessions, or finding public datasets. Triggers on "geo", "gene expression omnibus", "expression dataset", "RNA-seq dataset", "microarray dataset", "GSE", "GDS".

Workflow for ligand-receptor communication inference in single-cell or spatial data with sender-receiver summaries and cautious interpretation.

End-to-end EHR predictive modeling pipeline with PyHealth, covering dataset loading, task definition, model training, evaluation, calibration, and clinical interpretation.

> ESM2 protein language model for sequence scoring, embeddings, and plausibility checks. (2) Getting protein embeddings for clustering, (3) Filtering designs by sequence plausibility, (4) Zero-shot variant effect prediction, (5) Analyzing sequence-function relationships. For structure prediction, use chai1-structure-prediction or boltz-structure-prediction. For QC thresholds, use protein-design-qc.

Workflow for regulatory network inference, regulon scoring, perturbation-aware comparison, and network visualization.

Trim multiple sequence alignments using ClipKIT, trimAl, BMGE, Divvier, or HMMcleaner with mode selection guidance per downstream goal. Use when removing unreliable columns or contaminating residues before phylogenetic inference, HMM building, or selection analysis.

Amplicon sequence variant (ASV) inference from 16S rRNA or ITS amplicon sequencing using DADA2. Covers quality filtering, error learning, denoising, and chimera removal. Use when processing demultiplexed amplicon FASTQ files to generate an ASV table for downstream analysis.

Maps query single-cell data to reference atlases using scArches transfer learning with scVI and scANVI models. Transfers cell type labels without retraining on combined data. Use when annotating new single-cell datasets using pre-trained reference models.

Download large datasets from NCBI efficiently using EPost, history server, batching, rate limiting, and retry logic. Use when bulk-fetching tens of thousands of sequences, pulling all results of a large ESearch, designing reproducible pipelines, comparing E-utilities to NCBI Datasets v2 CLI, or implementing checksum-validated downloads. Encodes WebEnv TTL (~8h), EPost 200-ID limit, retmax caps, parallelization design, and integrity verification.

Integrate multiple scRNA-seq samples/batches using Harmony, scVI, Seurat anchors, and fastMNN. Remove technical variation while preserving biological differences. Use when integrating multiple scRNA-seq batches or datasets.

Run Bayesian phylogenetic analysis with MrBayes, BEAST2, RevBayes, and PhyloBayes including MCMC convergence diagnostics and model comparison. Use when needing posterior probability support, Bayesian model averaging, site-heterogeneous models for deep phylogenies, or formal model comparison via stepping-stone sampling.

End-to-end biomarker discovery workflow from expression data to validated biomarker panels. Covers feature selection with Boruta/LASSO, classifier training with nested CV, and SHAP interpretation. Use when building and validating diagnostic or prognostic biomarker signatures from omics data.

Pull pre-computed ortholog calls from public databases (OrthoDB, Ensembl Compara, OMA browser, eggNOG, PANTHER, KEGG Orthology, HomoloGene) via their REST APIs. Use when orthologs are already curated upstream, when the question is "what is the X ortholog of Y" rather than "how to infer orthology de novo", when batch-mapping gene IDs across species, or when comparing the resources for consensus calls. Encodes confidence-level semantics, 1:1 vs 1:many vs many:many, HomoloGene deprecation, and when to defect to de novo computation.

Reads, validates, and prepares CDISC SDTM and ADaM clinical trial data for analysis. Covers SDTM domain joins (DM, AE, EX, VS, LB, DS), ADaM architecture (ADSL, BDS, OCCDS, ADTTE) with traceability, treatment-emergent AE conventions, baseline derivation, SUPPQUAL/NSV handling, Define-XML 2.1, and Pinnacle 21 / CORE validation. Use when working with clinical trial datasets in CDISC SDTM/ADaM format, preparing analysis-ready data, or validating for regulatory submission.

Automated cell type annotation using reference-based methods including CellTypist, scPred, SingleR, and Azimuth for consistent, reproducible cell labeling. Use when automatically annotating cell types using reference datasets.

Trains and applies base-resolution deep learning models on ChIP-seq / ChIP-nexus / CUT&RUN data. Uses BPNet (Avsec 2021 Nat Genet 53:354; soft motif syntax from ChIP-nexus), chromBPNet (Pampari A et al 2025 Nat Genet; bias-factorized base-resolution profiles), EnFormer (Avsec 2021 Nat Methods 18:1196; 196 kb input, ~100 kb effective receptive field), DeepSEA (Zhou 2015; multi-task CNN), and JASPAR 2026 deep-learning collection (1259 BPNet ChIP models). Performs in silico mutagenesis for variant-effect prediction, DeepLIFT/Grad attribution, and TF-MoDISco motif discovery from attribution scores. Use when predicting variant effects on TF binding, discovering soft motif syntax / cooperativity, integrating ChIP-seq with sequence-only predictions, or applying precomputed JASPAR Deep Learning models to new variants.