fagonzalezo

Generates structured, rigorous experiment plans for machine learning research. Use this skill whenever a user wants to design, plan, or structure an ML experiment — including baseline comparisons, ablation studies, hyperparameter sweeps, architecture searches, data experiments, fine-tuning runs, or evaluation protocols. Trigger on phrases like "experiment plan", "research plan", "how should I test", "design an experiment", "ablation study", "benchmark this", "evaluate my model", "compare these approaches", or any request to systematically validate an ML idea or hypothesis. Also use for questions like "how do I know if X works better than Y" in an ML context. Produces a research-design plan (hypothesis, conditions, metrics, decision rules) — NOT a code-implementation plan. Implementation planning is a separate follow-up step, ideally run in plan mode after the experiment plan is accepted. Do NOT use for pure theory questions, math derivations, non-ML software engineering tasks, or for planning the code/scripts

Writes a structured report on ML experiment results, grounded in the experiment's DESIGN.md and the actual results files. Use this skill whenever a user wants to document, summarise, or communicate the outcome of a machine learning experiment — including phrases like "write a report", "summarise my results", "document what happened", "write up the experiment", "generate a summary of exp_N", "what did my experiment show", or "I finished running my experiment". Always use this skill when the user has completed (or partially completed) an experiment and wants a written record of findings. The report is written to experiments/exp_N/reports/summary.md. Do NOT use for planning new experiments (use ml-experiment-planner instead) or for pure data analysis without a write-up goal.

Guide users in applying the kdm-torch library (Kernel Density Matrices) to machine learning tasks. Make sure to use this skill whenever the user asks how to build a KDM model, use kdm-torch, or apply KDMs to classification, regression, density estimation, uncertainty quantification, or generative modeling. Trigger also when the user mentions kernel density matrices, KDMClassModel, KDMRegressModel, KDMDenEstModel, KDMLayer, init_kdm_layer, log_marginal, input log-likelihood, or OOD detection with KDMs.

Guide ML practitioners through analyzing, improving, and integrating embedding spaces using the EmbedKit library. Use this skill whenever the user has embedding arrays (NumPy/PyTorch), wants to assess embedding quality, diagnose geometric pathologies (hubness, anisotropy, distance concentration, intrinsic dimensionality mismatch), choose improvement strategies, apply contrastive learning to refine embeddings, or plug EmbedKit into their own ML workflows (HuggingFace, PyTorch training loops, scikit-learn pipelines, RAG systems, retrieval stacks). Trigger on: "embeddings", "hubness", "intrinsic dimension", "isotropy", "embedding space", "refine embeddings", "EmbedKit", sentence/image/word embedding quality, kNN quality, representation learning diagnostics, retrieval quality, semantic similarity, embedding geometry.

Build and maintain an LLM-curated personal knowledge base — the "LLM Wiki" pattern from Andrej Karpathy's April 2026 gist. Use this skill whenever the user wants to ingest a source (paper, article, transcript, PDF, notes) into a persistent compounding knowledge base, ask a question against accumulated notes, lint or audit such a base, or initialize a new one. Trigger on phrases like "add this to my wiki", "ingest this paper", "compile this into the knowledge base", "what does my wiki say about X", "lint the wiki", "build a knowledge base from these documents", "research notes", "second brain", "personal knowledge base", or any reference to LLM Wiki / OmegaWiki. Trigger even when the user does not say "wiki" — if they are accumulating sources over time and want them organized, this applies. The skill scales — sharded indexes, atomic pages, YAML frontmatter, and a bundled search script keep the wiki from becoming a context bottleneck at hundreds or thousands of pages.

Build and maintain an LLM-curated personal knowledge base — the "LLM Wiki" pattern from Andrej Karpathy's April 2026 gist. Use this skill whenever the user wants to ingest a source (paper, article, transcript, PDF, notes) into a persistent compounding knowledge base, ask a question against accumulated notes, lint or audit such a base, or initialize a new one. Trigger on phrases like "add this to my wiki", "ingest this paper", "compile this into the knowledge base", "what does my wiki say about X", "lint the wiki", "build a knowledge base from these documents", "research notes", "second brain", "personal knowledge base", or any reference to LLM Wiki / OmegaWiki. Trigger even when the user does not say "wiki" — if they are accumulating sources over time and want them organized, this applies. The skill scales — sharded indexes, atomic pages, YAML frontmatter, and a bundled search script keep the wiki from becoming a context bottleneck at hundreds or thousands of pages.