dsh-writing-shape

dsh-writing-shape is a writing-shape plugin that plugs into the DeepSeek Harness inference pipeline. In practice, when you call DeepSeek models to generate text, paragraph rhythm and sentence-level structure are hard to steer with any precision; output tends to drift toward either rambling or mechanical cadence. This plugin introduces the writing-shape dimension into the Harness reasoning flow, letting you exert finer-grained structural control over generated text without touching the underlying model weights.

Its core capability is to fold writing-shape parameters into the inference stage itself, so that Harness can honor explicit structural constraints while scheduling model output. In concrete terms, you can define distinct shape strategies for different writing scenarios—long-form review, multi-paragraph argumentation, tight abstracts—without resorting to heavy post-processing after generation.

It is aimed at developers building DeepSeek Harness inference workflows, especially teams that want text-structure controllability on the engineering side. Released under the MIT license, the plugin drops into existing Harness configurations with minimal integration overhead.