Best AI Agent Skills for Machine Learning and Data Science (2026)
ML code fails differently to application code. These SKILL.md skills encode reproducibility, notebook discipline, experiment tracking, and framework conventions.
ML code fails differently to application code. These SKILL.md skills encode reproducibility, notebook discipline, experiment tracking, and framework conventions.
Quick Answer: The best skills for ML and data science cover reproducibility (seeds, environment pinning, data versioning), notebook-to-module refactoring, experiment tracking conventions, pandas and PyTorch idioms, and data validation. All use the SKILL.md format and work across Claude Code, OpenClaw, Codex CLI, Cursor, and other compatible agents. Browse them at agensi.io/skills.
Why ML code needs different skills
Application code fails loudly. ML code fails quietly and produces a number anyway.
A data leak between train and test sets doesn't throw an exception, it just gives you a suspiciously good validation score. An unset random seed makes a result unreproducible without ever erroring. A SettingWithCopyWarning in pandas is a warning, not a failure, and the silently wrong dataframe flows downstream. A model that trains fine and scores well on a shuffled split can be worthless on a time series where the split leaked the future.
Generic coding skills check for bugs. ML work needs skills that check for the failure modes that don't raise.
Recommended skills
Csv Analyzer
by Kevin Cline
Automate data profiling with type detection, statistical analysis, and quality f…
Agent Framework
by Echo Rose
Agent Framework - A Premium AI Agent Skill
local-llm-troubleshooter
by LB Creations
Diagnose and fix broken local LLM stacks, GPU issues, and stalled model download…
Reproducibility skills
The single highest-value ML skill category, because irreproducibility is discovered late and costs the most.
A reproducibility skill should require: seeds set for Python random, NumPy, and the framework (torch.manual_seed, and CUDA seeds where relevant), environment pinned with exact versions rather than ranges, data version or hash recorded alongside results, and hyperparameters logged rather than left as literals in a cell.
It should also flag the things that silently break determinism — non-deterministic CUDA kernels, DataLoader with multiple workers and no seeded worker init, and dictionary iteration order assumptions.
truth-first (free) forces evidence-first verification of system state before acting on assumptions. It pairs well with ML work, where the temptation to report a result without re-running it is high.
Notebook discipline
Notebooks are where most ML work starts and where most of it should not stay.
A notebook skill should encode: no out-of-order execution assumptions, no hidden state carried between cells, imports at the top, and a clear boundary between exploration and anything that runs more than once.
The higher-value version is a notebook-to-module refactoring skill. It takes an exploratory notebook and extracts the reusable parts into importable functions with type hints and tests, leaving the notebook as a thin driver. That transformation is mechanical, tedious, and exactly what an agent should do.
Pandas and data manipulation
Pandas has a large surface area and a lot of ways to be subtly wrong.
Worth encoding in a skill: .loc and .iloc rather than chained indexing, explicit .copy() when a subset will be modified, pd.merge with an explicit how and validation of row counts after the join, and dtype checks after reading a CSV where pandas may have inferred object for a numeric column with one bad value.
Merge validation deserves its own rule. validate="one_to_one" or "one_to_many" turns a silent row explosion into an immediate error.
Browse data engineering skills on Agensi.
Framework conventions
PyTorch and TensorFlow have idioms an agent will mix if not told which you use.
For PyTorch: whether the project uses raw training loops, Lightning, or Accelerate. Device handling and whether .to(device) is centralised. model.train() and model.eval() discipline around evaluation. torch.no_grad() or inference_mode() on validation passes. Gradient accumulation and clipping conventions.
For scikit-learn: whether transformations live in a Pipeline rather than being applied ad hoc, which is the difference between a model you can deploy and one you can't.
Experiment tracking and MLOps
mlops-pipeline-generator generates ML pipeline scaffolding on Agensi.
An experiment tracking skill should state which tool the team uses — MLflow, Weights and Biases, or a plain results table — what gets logged as a matter of course (params, metrics, artifacts, git SHA, data version), and the run naming convention. The value is consistency: experiments that can't be compared because they logged different things are experiments you have to run again.
Data validation
Schema validation on inputs catches the class of bug where the pipeline runs successfully on the wrong data.
A validation skill should define expected schema, null tolerance per column, value ranges, and what happens on violation — fail loudly, or quarantine and continue. Whether you use Great Expectations, Pandera, or hand-rolled assertions, the rule should be that data entering a pipeline is checked rather than assumed.
Installing these skills
~/.claude/skills/
├── truth-first/
│ └── SKILL.md
├── code-reviewer/
│ └── SKILL.md
Claude Code reads from ~/.claude/skills/, OpenClaw from ~/.openclaw/skills/, Codex CLI from ~/.codex/skills/. Skills load at session start, so restart after adding one.
Full walkthrough: How to Install Skills in Claude Code.
Browse ML-compatible skills for any AI coding agent on Agensi.