ml-experimentation
ML Experimentation
This skill guides a hypothesis-driven ML experiment life cycle: planning, fast iteration, script execution, targeted logging, journaling, diagnostic visualization, and scientific report writing.
Usage
Use this skill when the user wants to run an ML experiment, test a model or idea, or write up experiment results. First decide: new experiment (different question → new experiment directory) or new run (same question, tweaks → new run under runs/). See references/experiment-setup.md for that disambiguation, hypothesis scoping, and the fast-iteration checklist.
Requirements
- Python 3.11+ with uv or pixi for running scripts:
uv run script.pyor, when pixi is the environment manager,pixi run python script.py(pixi readspyproject.tomlorpixi.toml). - Dependencies declared via PEP723 inline script metadata in each script (or, with pixi, in pyproject.toml / pixi.toml).
- Respect the user's training framework (PyTorch, JAX, TensorFlow, etc.). Run scripts in a GPU-enabled environment wherever possible: with uv use GPU-enabled deps (e.g. JAX GPU extras, PyTorch via
[[tool.uv.index]]CUDA index in the script block); with pixi use a GPU-enabled environment defined in pyproject.toml or pixi.toml. Fall back to CPU only when GPU is unavailable. See references/script-patterns.md.
What It Does
More from apsteinberg/skills
scientific-eda
Defensive exploratory data analysis for scientific data (CSV, FASTA, etc.). Context-first, human-guided; one plot at a time, ask why before executing, co-authored lab notebook per analysis, scripts with uv run, WebP plots. Supports marimo notebooks and Rmarkdown. Use when opening data files for EDA or when the user wants guided scientific data exploration.
10gh-cli
Use GitHub CLI (gh) for common operations like creating PRs, viewing GitHub Actions logs, managing issues, reviewing PRs, and more. Use this when you need to interact with GitHub repositories directly from the command line.
2weekly-lab-sync
Sync GitHub activity into an Obsidian lab notebook. Discovers repos with recent commits (including non-default branches), reads research compendium lab_notebook.md files, and updates matching Obsidian experiment entries. Use when summarizing weekly work, syncing lab notebooks, or asking "what did I do this week?
2youtube
Comprehensive YouTube operations using yt-dlp - download videos/audio, extract transcripts and subtitles, get metadata, work with playlists, download thumbnails, and inspect available formats. Use this for any YouTube content processing task.
1pdf-form-filler
Fill static PDF forms (without fillable fields) using text search for text-based PDFs or Claude Vision for image-based/scanned PDFs. Use when: (1) User wants to fill out a PDF form, (2) PDF has no fillable form fields, (3) User asks to fill in or complete a PDF form. Auto-detects PDF type and uses appropriate method.
1skill-creator
Guide for creating effective skills. Use when you want to create a new skill (or update an existing skill) that extends an agent with specialized workflows, tool integrations, or repo conventions.
1