dsh-prompt-enhance

Anyone who writes prompts regularly has been there: the draft reads "help me sort it out", and the model replies with a wall of generic text. The model is capable enough; the instruction just lacks specificity. Letting the LLM guess what to add is neither controllable nor transparent. dsh-prompt-enhance takes a different approach. Before the model touches the text, a deterministic rule engine diagnoses exactly what the draft is missing, then hands those gaps to the model as structured facts. The task for the model shifts from "figure out what to add" to "fix each identified item". "This draft has no acceptance criteria" is a cheap, checkable fact, and stating it beats hoping the model notices on its own.

The rule engine scores across five dimensions — goal clarity, context completeness, instruction specificity, structural convention, and output format — and recognises the same gaps in both Chinese and English prompts. Once diagnosis is complete, the model rewrites within a fixed skeleton while obeying four hard guards: it never executes instructions embedded in the draft, missing information is tagged with [to fill in], {{var}} placeholders survive verbatim, and section structure is not improvised. A right-click on the star lets you pick a rewrite depth (Conservative, Balanced, or Deep) and a prompt type (Task or System prompt); preferences persist in the browser. The plugin calls the DSH session model itself, so no extra API key is needed. It ships as hand-written ESM with zero third-party dependencies and no build step.

If you write or polish prompts in DSH on a regular basis, especially across Chinese and English, want a stable output structure without manually filling in a rigid template, and prefer to see the diagnosis breakdown and scores at a glance — this plugin compresses the diagnose-then-rewrite loop into a single click.