Compose method atoms into a validated training recipe JSON.
npx skills add https://github.com/chenghaoYang/auto-coder-trainer --skill compose-chenghaoyang
Please help me install this Agent Skill. Skill: compose Source: https://github.com/chenghaoYang/auto-coder-trainer/tree/main/skills/compose Command: npx skills add https://github.com/chenghaoYang/auto-coder-trainer --skill compose-chenghaoyang
This skill enables teams to quickly assemble modular method atoms into a complete training recipe, reducing setup time and improving reproducibility.
Create a new training recipe by selecting atoms and saving it to recipes/examples/compose.recipe.json.
High-intent search queries and answers about installing and using this skill.
Assembling method atoms into a training recipe involves loading atoms from a registry, selecting them by name or description, and merging them into a Recipe IR with model, dataset, and trainer configurations.
A Recipe IR is an intermediate representation that standardizes experiment configuration by combining model, dataset, trainer, and budget components into a single coherent structure for validation and execution.
Yes, training recipes are validated against recipes/schema/recipe.schema.json, which enforces required fields including id, name, model, dataset, and trainer to ensure configuration integrity before execution.
You load method atoms from recipes/registry/method_atoms.json and select specific atoms by matching their name or description to align with your workflow intent before merging them into a recipe.
Yes, after validation the assembled training recipe is saved to recipes/examples/<name>.recipe.json, creating a persistent configuration file ready for downstream execution and reproducibility.
The required fields for composing a training recipe are id, name, model, dataset, and trainer, which are strictly enforced during schema validation to ensure structural completeness.
chenghaoYangCollect coding agent training papers and methods into a centralized registry.
thinking-machines-labScaffold a new training recipe directory with boilerplate in tinker_cookbook/recipes.
TeoSlayerDeploy a four-agent ML training pipeline coordinating data preparation, training, evaluation, and serving.
Devsoul2026Track ML experiments and metadata with Weights & Biases.
