---

name: module-authoring description: "Author and debug custom Sonnet modules, lazy variables, composition wrappers, and module build contracts." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

Sonnet Module Authoring

Use this sub-skill for custom snt.Module classes, @snt.once lazy variable creation, variable and submodule inspection, name scopes, snt.Sequential, snt.Deferred, snt.BatchApply, and snt.build shape checks.

Start here

1. Read references/module-patterns.md for concrete authoring patterns.
2. Read references/api-reference.md for verified signatures and contracts.
3. Read references/troubleshooting.md for construction, first-call, and composition failures.
4. Run scripts/module_contract_smoke.py for a tiny executable lazy-variable and optional BatchApply check.

Boundaries

• Built-in layers and snt.nets: ../layers-and-nets/SKILL.md.
• Optimizers and training loops: ../training-and-optimization/SKILL.md.
• Functional init/apply APIs: ../functional-transforms/SKILL.md.
• Checkpoint/SavedModel/distribution: ../serialization-and-distribution/SKILL.md.

Minimum module contract

• Subclass snt.Module and call super().__init__(name=name) in __init__.
• Put input-dependent variable creation in a side-effect-only @snt.once method that returns None.
• Call/build the module before inspecting variables or trainable_variables.
• Use explicit subclasses for branching or extra call arguments; use Sequential only for plain chains.
• Use Deferred when one module's constructor depends on another module after first call, and BatchApply for applying a 2-D module over extra leading dimensions.