dsh-prompt-optimizer

Writing a prompt the model actually executes correctly is often harder than writing the code itself: vague roles, missing constraints, unclear output format—and after three rewrites it still is not right. dsh-prompt-optimizer drops a single ✨ button right next to the send bar in DeepSeek Harness. Click it and the current draft, together with recent conversation context, is sent to the model; a five-dimension diagnosis and a fully rewritten prompt stream back in real time, all without leaving the chat window.

Under the hood the plugin auto-selects between two strategies. When context is thin it normalises the draft into a role–task–constraint–output structure; when context is rich it first distils the real intent, then polishes the wording while preserving your original framing—no template stuffing, no re-asking what the conversation has already established. Every output passes a fidelity self-check: semantic equivalence is the hard floor, inferences are flagged with an explicit caveat, and drift is caught before the result is shown. A lightweight memory chain lets you iterate naturally—edit the optimised draft, hit optimise again, and the plugin carries forward confirmed decisions, adjusting only around what changed. Streaming display, a timeout watchdog, cancel-without-loss, slash-command prefix protection, and per-session panel isolation keep the experience reliable under everyday high-frequency use.

It is built for anyone who talks to models often inside DeepSeek Harness and wants prompt quality to keep improving without extra tabs or copy-paste rituals: drafting technical specs, debugging complex multi-step instructions, standardising prompt templates for a team, or simply refusing to spend another ten minutes tweaking wording. Install it, the button appears the moment there is text in the input box, and the rest is up to you.