new-recipe-template

name	new-recipe-template
description	Create a new recipe scaffold under recipes/ with standard config, engine, dataset, model, README, and recipe-local test structure while keeping task-specific implementation intentionally minimal.

New Recipe Template

Goal

Create a clean recipe scaffold under recipes/<recipe_name>/:

• keep experiment-specific logic inside the recipe;
• create standard config, engine, dataset, model, and README files;
• include recipe-local tests/test_structure.py and tests/test_smoke.py when requested;
• leave task-specific data/model logic explicit and minimal until requirements are known;
• avoid adding repo-wide abstractions.

Required Inputs

Collect these before writing files:

• recipe name in snake_case;
• short task summary;
• config filename, default train;
• whether to include recipe-local tests, default true;
• closest existing recipe to use as a reference, if any.

Ask once in a compact message when any required input is missing.

Workflow

1. Validate Naming And Scope

Confirm:

• recipe folder is snake_case;
• default engine class is <RecipeNamePascalCase>Engine;
• config class is <RecipeNamePascalCase>Config;
• no existing recipe path will be overwritten.

2. Create Standard Layout

Create:

recipes/<recipe_name>/
├── README.md
├── __init__.py
├── configs/
│   ├── __init__.py
│   ├── schema.py
│   └── <config_name>.yaml
├── dataset/
│   └── __init__.py
├── engine/
│   ├── __init__.py
│   └── <recipe_name>_engine.py
├── model/
│   └── __init__.py
└── tests/
    ├── test_structure.py
    └── test_smoke.py

Use tests/templates/test_structure.py.template and tests/templates/test_smoke.py.template when creating tests.

Read references/scaffold_rules.md before drafting files.

3. Keep Stubs Honest

Use explicit NotImplementedError for task-specific engine methods when the real implementation is unknown. Do not guess dataset/model behavior.

The README and config should describe the provided task summary and clearly mark remaining implementation work.

4. Validate The Scaffold

Run:

python3 -m compileall recipes/<recipe_name>
pytest recipes/<recipe_name>/tests/test_structure.py -q

Run smoke only after the recipe has real dataset/model/engine behavior.

Validation

Soft Validation

Review the scaffold without running tests:

• names and imports match the recipe name;
• config schema validates the YAML shape;
• engine is registered and names match;
• dataset and model remain minimal unless task-specific behavior is known;
• tests were created from current repo templates when requested;
• no mvp_engine/ changes were introduced.

Hard Validation

If recipe-local tests exist, optionally copy and adapt references/asserts.py into:

recipes/<recipe>/tests/skills/new-recipe-template/asserts.py

Run:

python3 -m compileall recipes/<recipe>
pytest recipes/<recipe>/tests/test_structure.py -q

Do not require tests/test_smoke.py to pass until real training behavior is implemented.

Output

• State recipe path created.
• State recipe name, config name, and whether tests were included.
• List files created.
• State placeholders or NotImplementedError methods that still need real implementation.
• State validation commands and results.

Read On Demand

• references/scaffold_rules.md: standard file contents and scaffold choices.
• references/example.md: sample prompt flow and expected tree.