dsh-prompt-enchant

Prompt quality still caps model output, yet we think in loose, colloquial fragments. dsh-prompt-enchant slots a sparkle-star magic-wand button next to the DeepSeek Harness web input so you can type whatever is in your head, tap the star, and a single independent LLM call rewrites the messy draft into a clearer, more executable instruction that fills back into the input for your review before sending. The enhancement never touches the agent system prompt, and if the result is off you simply discard it.

Two persona modes are built in. General mode covers everyday conversation and workflow tasks. Design mode is tuned for AI image and video creation, auto-surfacing professional terminology around subject, style, composition, lighting, camera motion, and first/last-frame parameters; elements the model cannot confidently infer are routed through the official confirmation card for per-item selection. Working in General mode but leaning creative? A suggestion bar appears above the input with one-click switching and automatic re-enhancement. Safety guards include verbatim preservation of @file references, untouched /plan-style command declarations, plain-text output with injection defense, loopback-only binding, cross-origin rejection, and a concurrency cap, so the enhancement never disrupts your existing workflow.

This plugin suits DeepSeek Harness web users who want fewer long prompt-writing sessions without sacrificing precision, and especially creators iterating on text-to-image, text-to-video, or interactive editing tasks. No extra API key is needed; the enhancement call rides on DSH's built-in llm service and all data stays inside the local process.