dsh-prompt-polish

dsh-prompt-polish

English · 中文

You type "帮我写个周报" into DSH, hit Enter, and get a generic answer. You know the model could do it well — if only the prompt had a role, a goal, the data you have, and an output format. Writing all that out every time is the boring part.

This plugin adds a ✨ 润色 (Polish) button right next to the model selector in the DSH composer. Click it, and your rough draft is rewritten in-place — streamed into a panel — into a structured prompt. You decide what happens next: write it back into the composer (default, nothing is sent), send it directly, or just copy it.

Nothing about the polish is silent: you see the result, the model that produced it, how long it took, and what it cost in tokens.

What it actually does

1. Rewrites your draft, you keep control. The optimized prompt lands in a panel, not in your composer. Three buttons decide its fate:

• 替换草稿 (Replace draft) — writes the result into the composer. Nothing is sent; you still read and edit it. This is the default path and the whole point: the AI prepared it, you pull the trigger.
• 发送草稿 (Send draft) — writes it back AND sends, one click fewer.
• 复制草稿 (Copy) — clipboard only.

2. Knows the difference between a system prompt and a user prompt. The panel top row is not a style picker — it picks which template family rewrites your text: 系统提示词 (you are writing an AI's instructions), 用户提示词 (you are asking for work), or 迭代优化 (revise the previous result: make it shorter, add an example, tone it down for newbies — the revision note is required, and the button tells you so if you skip it).

3. Shows what each polish cost. Every finished run reports: model, wall time, result length, and token usage (120 in + 45 out = 165 tokens) when the provider reports it. When the provider does not, it says not-provided — it never invents a zero.

4. Keeps a local history you explicitly opt into. The first time a run finishes, a card asks: history stays on this device, contains inputs/results/models/tokens, and can be wiped anytime. 开启历史 saves; 不保存 keeps nothing and never nags again (re-enable any time from the History view). With consent, each run becomes a version, and revising a version chains them, so you can:

• re-read any run's original input, output, and token cost;
• open 对比版本链: original | selected | latest side by side, plus a local (deterministic, no-model-call) structural analysis — lines added/removed, and whether goal/background/constraints/output-format coverage appeared or disappeared;
• delete one version, delete a whole chain, or clear everything.

5. Extracts variables. One click asks the model which spans of the finished prompt are the changeable parts (time range, audience…) and rewrites them into {{变量}} slots — a reusable template you write back.

When it is worth using (and when it is not)

Worth it: short or vague drafts; recurring task types (reports, analyses, code requests); when you keep re-explaining what you want.

Skip it: prompts that are already structured (role + context + constraints + format); code snippets where you do not want the model rewording you; anything where you do not want an extra model call — the polish is a model call, and on a slow model you will wait.

Honest limitations: the result is only as good as the template and model behind it — it adds structure but cannot know facts you forgot to include; iteration is one revision per call, not a chat; and every click costs tokens (typically a few hundred, shown before you commit).

Two real examples: what changes is the completeness of the task

The short version first: polish never rewrites your wording — it writes out the goal, the basis, the constraints and the output format that lived only in your head. Both examples below are real runs from this plugin's own usage.

Example 1: one line of spoken-style feedback → seven executable requirements

Typed (7 short sentences, all attitude):

你不要僵硬，实际都是以具体内容为准的。你的目的在于梳理该章节的内容，以思维导图的形式呈现。 ("Don't be rigid — go by the actual content. Organize this chapter as a mind map.")

The trouble: "don't be rigid" is an attitude, not a standard. What counts as rigid, what to go by, how many branches, what the output looks like — the model can only guess, and may guess wrong.

After one ✨ click, the model received:

请以我提供的章节具体内容为唯一依据，梳理该章节的内容结构，并输出为思维导图。要求如下： 1. 不要套用固定模板、预设框架或僵化层级；章节实际讲了什么，就按什么逻辑拆分。 2. 中心主题：使用章节标题或能概括全章的核心主题。 3. 一级分支：按主要部分/核心论点/关键问题划分，通常 3—7 个，不强行凑数。 4. 下级分支：按分类、因果、流程、对比等实际逻辑展开，简单到二级、复杂到三四级。 5. 节点表述：简洁名词短语，保留原文术语，不添加、不遗漏。 6. 输出形式：Markdown 嵌套列表，体现父子层级。 7. 多种合理梳理时，选最贴合原文结构、最便于理解的一种。

What changed: one attitude became seven checkable standards — a single source of truth, structure-follows-content, branch counts, depth limits, node phrasing, output format, and how to choose among competing readings. Every place you feared "rigid" now maps to a hard requirement.

Example 2: "take a look" → a complete review brief

Typed (9 characters of intent; no criteria, no output spec):

我是想让你从产品经理的角度，看看我们这个作品设计的咋样，还有什么需要改进的 ("From a PM's perspective, look at our design — how is it, what should improve?")

This names a role (PM) but not what to review against, not what the deliverable looks like. The likely reply is polite filler: "overall good, polish the details."

After one ✨ click, the model received:

请你以产品经理的视角，评审我提供的作品设计并给出改进建议。作品信息如下（请按实际情况补充）：作品名称/类型、目标用户群体、要解决的核心需求或痛点、核心功能与主要使用流程、设计材料。请重点从以下维度分析：1. 目标用户与核心需求是否明确、痛点是否真实；2. 功能与场景是否匹配、有无冗余或缺失；3. 流程与交互是否顺畅、上手成本；4. 与同类产品的差异和优势；5. 可行性与商业价值。输出要求：先整体评价（优点+最关键问题），再按高/中/低优先级列可落地建议（每条带理由和预期效果），信息不足先列需要补充的内容，不要凭空假设。

What changed: three moves — it asks you for the missing product info instead of inventing it, locks "take a look" into five judgeable dimensions, and fixes the output order: overall verdict → prioritized actionable suggestions → request materials first if inputs are missing. The exits for generic praise are closed, one by one.

The pattern in both: polish supplies the premises your original message never stated — the basis, the standards, the boundaries, the deliverable shape. The fuller those premises, the closer the result to what you meant. That is what the templates do, and what appending "take it seriously" never will.

Where it came from

The optimization engine is a faithful port of linshenkx/prompt-optimizer (AGPL-3.0): the three optimizationMode families, the template texts (kept verbatim, 10/10 byte-checked against upstream), the mustache rendering subset, the template registry with bidirectional field/reference validation, and the variable-extraction flow. The composer button, panel UI, history/version chain, comparison, and token reporting are new in this repo.

Install

DSH (desktop or web profile) + Node.js 18+.

The plugin is on npm as @benrong/dsh-prompt-polish, so on a CLI-managed profile (e.g. web) the stock command works:

dsh plugin --profile web add @benrong/dsh-prompt-polish

On the desktop profile the Electron app owns the profile (dsh plugin --profile desktop … is refused), so use the script below instead.

git clone https://github.com/benrong2048-boop/dsh-prompt-polish.git
cd dsh-prompt-polish
pwsh -File install-plugin.ps1 -ProfileName desktop   # or -ProfileName web

The script copies the package into %USERPROFILE%.dsh\profiles<profile>\node_modules, syntax-checks it, and appends the plugin row to that profile's cordis.patch.yml (it prints the exact lines it writes). Then restart DSH — a full restart; the desktop app has no reload key, and Ctrl+R does nothing.

Manual install, if you would rather see everything it touches:

1. Copy this folder to %USERPROFILE%.dsh\profiles<profile>\node_modules@benrong\dsh-prompt-polish.
2. Append to %USERPROFILE%.dsh\profiles<profile>\cordis.patch.yml:

- insert:
    - id: prompt-polish
      name: "@benrong/dsh-prompt-polish"

3. Restart DSH.

Uninstall: delete the package folder and the insert block above, restart DSH.

Cost and privacy

• Model calls: one streaming call per polish click, one per 迭代优化 revision, one per 提取变量. Nothing runs in the background; opening the panel costs nothing. Token counts are shown per run and per version chain.
• Privacy: history (if you consent) lives in the browser localStorage on your machine. No telemetry, no uploads, no account. Deleting history is real deletion.
• Disk: the installed package is ~90 KB (seven lib files plus the loader patch); localStorage history is capped at 50 entries with 4,000-character clipping per field.

Tests

node mustache.test.mjs    # rendering engine semantics (34)
node smoke.mjs            # template layer (219)
node contract.test.mjs    # browser half, real useEffect + fake streaming fetch (117)
node host.test.mjs        # host half, three routes + failure paths (164)
node compare-upstream.mjs # audit: template texts vs upstream (needs $env:UP=<upstream checkout>)

534 assertions, all green at publish time.

Submit to dshmarket

A ready-to-paste catalog entry lives in market/plugins-entry.json, with a step-by-step guide in market/SUBMITTING.zh.md (publish to npm, then PR the entry to the awesome-dsh-plugin catalog).

License

AGPL-3.0-or-later. See LICENSE. Upstream template texts and flows come from linshenkx/prompt-optimizer, also AGPL-3.0 — the license is inherited, and their authorship is credited.