Skip to content
Back to search

train-recipe

SkillWorks with Claude CodeGitHub

Stage-by-stage, gate-verified process for training, debugging and improving deep learning and tabular ML models, based on Karpathy's recipe, with an experiment log

Safety scan

A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.

Checked today
Badge for your README
findsafeskills safety rating

Markdown

[![findsafeskills safety rating](/api/listings/misobalogh--train-recipe/badge.svg)](/listings/misobalogh--train-recipe)

HTML

<a href="/listings/misobalogh--train-recipe"><img src="/api/listings/misobalogh--train-recipe/badge.svg" alt="findsafeskills safety rating" /></a>

Install

i

Install instructions from the README

As written by the repo's author. findsafeskills has not verified these steps.

The skill is a plain `SKILL.md` folder (`skills/train-recipe/`) in the open Agent Skills format, so any compatible agent can use it. Pick one of the options below.

Any agent: `npx skills`

The [skills CLI](https://github.com/vercel-labs/skills) detects your installed agents and copies the skill into the right folder:

npx skills add misobalogh/train-recipe          # current project
npx skills add misobalogh/train-recipe -g       # all projects
npx skills add misobalogh/train-recipe -a codex # one specific agent

Claude Code: plugin marketplace

The plugin is listed in the [misobalogh/claude-plugins](https://github.com/misobalogh/claude-plugins) marketplace and updates from there:

/plugin marketplace add misobalogh/claude-plugins
/plugin install train-recipe@misobalogh

The shell equivalents are `claude plugin marketplace add ...` and `claude plugin install ...`. To invoke the skill explicitly, use `/train-recipe:train-recipe`.

Codex

Ask the built-in installer: `$skill-installer install the train-recipe skill from https://github.com/misobalogh/train-recipe`. Or use `npx skills` or copy the folder manually (see below). Restart Codex if the skill does not show up.

Manual copy

git clone https://github.com/misobalogh/train-recipe.git
cp -r train-recipe/skills/train-recipe <skills-dir>/

| Agent | Project `<skills-dir>` | Global `<skills-dir>` | |-------|------------------------|-----------------------| | Claude Code | `.claude/skills/` | `~/.claude/skills/` | | Codex | `.agents/skills/` | `~/.agents/skills/` | | Cursor | `.agents/skills/` | `~/.cursor/skills/` | | Gemini CLI | `.agents/skills/` | `~/.gemini/skills/` | | GitHub Copilot | `.agents/skills/` | `~/.copilot/skills/` | | OpenCode | `.agents/skills/` | `~/.config/opencode/skills/` | | Windsurf | `.windsurf/skills/` | `~/.codeium/windsurf/skills/` |

Agents change these paths from time to time. If one does not pick the skill up, check its docs.

Agents without skill support

Copy the `skills/train-recipe/` folder into your project and add this line to your `AGENTS.md` (or the agent's rules file, e.g. `CLAUDE.md`, `GEMINI.md`, `.cursorrules`):

When training, debugging, tuning or auditing an ML/DL model, read and follow
`skills/train-recipe/SKILL.md` and the reference file it points to.

What was scanned

The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.

.claude-plugin/plugin.json · Manifest · 13 lines
  1. {
  2. "name": "train-recipe",
  3. "description": "Stage-by-stage, gate-verified process for training, debugging and improving deep learning and tabular ML models, based on Karpathy's recipe, with an experiment log",
  4. "version": "0.2.0",
  5. "author": {
  6. "name": "Michal Balogh",
  7. "url": "https://github.com/misobalogh"
  8. },
  9. "homepage": "https://github.com/misobalogh/train-recipe",
  10. "repository": "https://github.com/misobalogh/train-recipe",
  11. "license": "MIT",
  12. "keywords": ["machine-learning", "deep-learning", "training", "debugging", "tabular", "karpathy"]
  13. }
README.md · README · 135 lines
  1. # train-recipe
  2. An agent skill for Claude Code, Codex, Cursor, Gemini CLI, GitHub Copilot, OpenCode and any other
  3. agent that supports the open [Agent Skills](https://agentskills.io) format. It makes the agent train, debug and improve ML models the way
  4. Andrej Karpathy describes in
  5. [A Recipe for Training Neural Networks](https://karpathy.github.io/2019/04/25/recipe/) (2019),
  6. updated for current practice and extended to classic/tabular ML. Most ML skills are a collection of
  7. tips. This one **enforces the process**: the agent works through six stages in order, checks each
  8. stage's gate with a real measurement before moving on, goes back when an earlier gate was never
  9. verified (even if you asked about tuning), and keeps an experiment log (`RECIPE_LOG.md`) in your
  10. project with one line per experiment: hypothesis -> change -> result -> keep/revert.
  11. It is process-only: no scripts. Every check is described precisely enough (what to measure, expected
  12. value, what a failure means) for the agent to write it inside your codebase, in whatever framework
  13. you use.
  14. All credit for the original method goes to Andrej Karpathy. This skill paraphrases and extends his
  15. post; read the original.
  16. ## The stages
  17. 1. Become one with the data
  18. 2. End-to-end skeleton + dumb baselines
  19. 3. Overfit
  20. 4. Regularize
  21. 5. Tune
  22. 6. Squeeze out the juice
  23. ## What is modernized vs. the 2019 post
  24. - Pretrained weights (timm / Hugging Face / torchvision) as the default starting point.
  25. - AdamW, LR range test, warmup + cosine added later as a separate experiment.
  26. - Mixup / cutmix, label smoothing and EMA weights in the regularization ladder.
  27. - Silent bugs of current stacks: mixed precision / bf16, BatchNorm with tiny batches, GPU
  28. nondeterminism, `torch.compile`, gradient clipping.
  29. - Per-example gradient dependency test, also along the time axis for causal models.
  30. - Random or TPE search with multiple seeds; calibration, distillation and TTA at the end.
  31. - A full tabular / classic ML branch: leakage review, time and group splits, the
  32. constant -> linear -> GBDT baseline ladder, fold variance, and when a neural net is worth it.
  33. - An explicit experiment log and three working modes (new project, audit, triage).
  34. ## Install
  35. The skill is a plain `SKILL.md` folder (`skills/train-recipe/`) in the open Agent Skills
  36. format, so any compatible agent can use it. Pick one of the options below.
  37. ### Any agent: `npx skills`
  38. The [skills CLI](https://github.com/vercel-labs/skills) detects your installed agents and copies the
  39. skill into the right folder:
  40. ```bash
  41. npx skills add misobalogh/train-recipe # current project
  42. npx skills add misobalogh/train-recipe -g # all projects
  43. npx skills add misobalogh/train-recipe -a codex # one specific agent
  44. ```
  45. ### Claude Code: plugin marketplace
  46. The plugin is listed in the [misobalogh/claude-plugins](https://github.com/misobalogh/claude-plugins)
  47. marketplace and updates from there:
  48. ```
  49. /plugin marketplace add misobalogh/claude-plugins
  50. /plugin install train-recipe@misobalogh
  51. ```
  52. The shell equivalents are `claude plugin marketplace add ...` and `claude plugin install ...`.
  53. To invoke the skill explicitly, use `/train-recipe:train-recipe`.
  54. ### Codex
  55. Ask the built-in installer: `$skill-installer install the train-recipe skill from
  56. https://github.com/misobalogh/train-recipe`. Or use `npx skills` or copy the folder
  57. manually (see below). Restart Codex if the skill does not show up.
  58. ### Manual copy
  59. ```bash
  60. git clone https://github.com/misobalogh/train-recipe.git
  61. cp -r train-recipe/skills/train-recipe <skills-dir>/
  62. ```
  63. | Agent | Project `<skills-dir>` | Global `<skills-dir>` |
  64. |-------|------------------------|-----------------------|
  65. | Claude Code | `.claude/skills/` | `~/.claude/skills/` |
  66. | Codex | `.agents/skills/` | `~/.agents/skills/` |
  67. | Cursor | `.agents/skills/` | `~/.cursor/skills/` |
  68. | Gemini CLI | `.agents/skills/` | `~/.gemini/skills/` |
  69. | GitHub Copilot | `.agents/skills/` | `~/.copilot/skills/` |
  70. | OpenCode | `.agents/skills/` | `~/.config/opencode/skills/` |
  71. | Windsurf | `.windsurf/skills/` | `~/.codeium/windsurf/skills/` |
  72. Agents change these paths from time to time. If one does not pick the skill up, check its docs.
  73. ### Agents without skill support
  74. Copy the `skills/train-recipe/` folder into your project and add this line to your
  75. `AGENTS.md` (or the agent's rules file, e.g. `CLAUDE.md`, `GEMINI.md`, `.cursorrules`):
  76. ```markdown
  77. When training, debugging, tuning or auditing an ML/DL model, read and follow
  78. `skills/train-recipe/SKILL.md` and the reference file it points to.
  79. ```
  80. ## Example prompts
  81. New project:
  82. - "Set up training for an image classifier on the data in `data/`, follow the training recipe."
  83. - "I want to predict churn from `customers.parquet`. Start from scratch."
  84. Audit existing code:
  85. - "Audit my training code in `src/train.py` against the recipe and tell me which gates are missing."
  86. - "Review this LightGBM pipeline for leakage and missing baselines."
  87. "Model isn't learning" triage:
  88. - "My val loss has been stuck at 0.69 since epoch 1. What's wrong?"
  89. - "Loss goes to NaN after a few hundred steps with bf16."
  90. - "The model is 99% accurate on validation, that seems too good."
  91. ## Repository layout
  92. ```
  93. .claude-plugin/
  94. plugin.json # Claude Code plugin manifest (other agents ignore it)
  95. skills/train-recipe/ # the portable Agent Skills folder
  96. SKILL.md # shared backbone: rules, modes, stages, gates, common checks
  97. references/
  98. deep-learning.md # DL-specific checks per stage
  99. tabular-ml.md # tabular / classic ML checks per stage
  100. ```
  101. ## License
  102. [MIT](LICENSE), Copyright (c) 2026 Michal Balogh.

Version history

  • v0.2.011/19/2026

Ratings and comments

Sign in to rate this skill.

Sign in

Free to use. If it’s helped, pass it on.