Authoring Dynamic Code Generation Skills

Descriptor Surface

The source descriptor for this family is dcgskill.md. During preparation, the subsystem reads the descriptor through the generic skill document parser and then extracts a small set of sections: prompt guidance, argument description, and LLM mode. The recognized aliases are deliberately narrow. Prompt supplies the behavioral guidance. Argument, Input, or Parameters can supply a descriptive explanation of the expected input. LLM Mode, LLM-Mode, or Mode influence how the subsystem normalizes the execution tier.

The important limitation is that the argument section does not redefine the runtime argument key. The subsystem always treats the operative argument name as input. The descriptor may describe that input more precisely, but it does not rename it. If a caller passes structured arguments explicitly, the subsystem looks for args.input; otherwise it falls back to the raw prompt text.

The LLM mode field is also narrower than a casual reading might suggest. Internally, values that mention deep or code are normalized to code. Everything else becomes fast. The field therefore influences the runtime tier label rather than defining an arbitrary execution strategy vocabulary.

Canonical Layout

The valid layout is descriptor-only. In this form, the subsystem prepares its dynamic decision executor and lets the LLM decide at runtime whether the request should be answered directly or through temporary code.

This layout is operational. The runtime does not load executable modules from the skill folder. Authoring should therefore focus on descriptor guidance quality and clear input expectations in dcgskill.md.

Minimal Descriptor Example

A minimal descriptor for the default dynamic path may look as follows.

# Big Integer Multiplier

## Prompt
Decide whether to answer directly or to generate JavaScript that computes an
exact result. Prefer code when exact arithmetic is safer than free-form text.

## Argument
Primary natural-language instruction or text payload.

## LLM Mode
fast

This is enough for the subsystem to record prompt guidance, a human-readable argument description, and the normalized tier label before executing the standard decision path.