提问引导 · dsh-HelpMe-To-Ask

简体中文 · English

仓库：github.com/DoreiHime/dsh-help-me-to-ask

---

简体中文

你好呀，谢谢你点进来看。

做这个小插件的原因其实挺朴素的：我在 DSH 里干活的时候，经常一开始只说一句“帮我搞一下这个”，等模型吭哧吭哧做完，才发现方向不太对，又要推倒重来。后来就想，要是有个东西能在动手之前先轻轻问一句“你是想要 A 还是 B 呀”，大概能少很多来回。

所以有了它：回答之前，先帮你把话问清楚，再把你的回答整理成一张小小的「任务说明」（也可以理解成开工前的小单子），你确认之后才交给模型。就这些，没有别的了。

它会做什么

新建对话时，如果你说的话比较笼统，它会先弹一张卡片：

• 先复述一句“我理解你想做的是……”，让你看看有没有跑偏；
• 只问真正缺的东西：目标、交付物、范围限制、验收标准、有没有什么要特别注意的；
• 每个问题都尽量给出从你原话里推出来的具体选项，不太会问那种“请描述你的需求”的空话；
• 实在不想一项项选，可以点“一键明确目标”，让它自己补全。

你回答完（或者点了跳过），它会整理成一张很短的任务说明，大概长这样：

要做成什么 / 最后给你什么 / 限制（能改哪些、用什么）/ 怎样算做好 / 明确不碰的 / 没问到、先这样定

看一眼，确认、改一改、或者直接跳过。确认之后，模型就知道自己到底要做什么、做到什么程度算完成。

我不想被问的时候

输入框左边有个「提问引导」的小开关，点齿轮可以：

• 选什么时候问：只在需求模糊时问 / 复杂任务也问 / 每个新对话都问；
• 本会话跳过（只跳过当前这个对话，别的对话不受影响）；
• 调最多追问几轮（1–10，默认 3）；
• 关掉“开始前先看任务说明”，直接开工。

它也不会一直追问下去。没有新的信息可问了，它会停，把剩下的细节写成“假设”，或者留给你和模型在对话里慢慢补。

设置里还能改什么

设置 → 提问引导：

• 总开关：不想用了随时关掉；
• 提问时机：同上；
• 后续目标变化时提示更新任务说明：只在你明确说“换个 / 改成 / 重新来”这类话时才会出现，不会每轮都打扰；
• 开始前先看任务说明；
• 最多追问轮数；
• 大脑模型：插件自己分析需求时用的模型。默认自动选 DSH 里最新的 DeepSeek 模型 + Max 推理；也可以自己选客户端里任意可用的模型和推理档位。

怎么装

如果是从插件市场看到这个页面，点安装就好。

想手动装的话：

dsh plugin add dsh-help-me-to-ask

或者下载 release 里的 tgz，用本地安装。

装好之后重启一下 DSH（或者刷新页面），输入框左边就能看到它了。

一些说明

• 插件只在你新建对话的第一条消息，以及后面出现明显“改口”信号时才会出来；不会每轮都插话。
• 它不碰你的模型和推理档位设置。以前我做过一版会自动帮人切模型、切推理，后来发现身边几乎没人真的想为 DSH 选推理等级，就把它去掉了，只留下“问清楚”这件事。
• 正在执行任务的子智能体不会用到这个插件。
• 分析用的模型是你自己在设置里选的那个，调用方式和普通对话一样，数据不会额外传到别的地方。
• 会话里的任务说明只存在你自己电脑上：$DSH_HOME/dsh-help-me-to-ask-sessions.json。
• 它也有点笨的时候，如果问得不对、或者你觉得哪句话说得不清楚，欢迎告诉我，我慢慢改。

更新日志

见 CHANGELOG.md。

谢谢

写这个插件的时候被 DSH 的社区帮了很多，也谢谢愿意试用它的你。

如果有什么想跟我说的，可以发邮件到 doreihime.cn@gmail.com，我看到会回。

MIT License。

---

English

Hi, thank you for stopping by.

The idea behind this little plugin is pretty simple: when I work in DSH, I often start with something vague like “help me fix this”, and only after the model has done a whole round of work do I realize the direction was off. So I thought: what if something asked me one gentle question before it starts — “did you mean A or B?” — that would save a lot of back-and-forth.

That’s all it does. Before answering, it helps you state your request clearly, turns your answers into a tiny task sheet, and waits for your confirmation before handing it to the model.

What it does

When a new conversation starts with a vague message, it shows a card first:

• It repeats back “here’s what I understand you want…”, so you can catch a misunderstanding early;
• It asks only for what’s genuinely missing: goal, deliverable, constraints, acceptance criteria, anything special to watch out for;
• Each question offers concrete options derived from your own words, instead of empty prompts like “please describe your requirements”;
• If you don’t feel like picking each item, press One-click brief and it fills in the gaps itself.

After you answer (or skip), it produces a very short task sheet:

What to make / What you get / Limits (what may be touched, what to use) / How we know it’s done / What not to touch / Guesses for gaps

Check it, confirm, modify, or skip. Once confirmed, the model knows exactly what to do and what “done” means.

When it stays quiet

There’s a small Ask Guide switch on the left of the input box. Click the gear to:

• choose when it asks: only when the request is vague / also for complex tasks / every new conversation;
• skip in this conversation (only the current one, other conversations are unaffected);
• set the maximum number of follow-up rounds (1–10, default 3);
• turn off “review the task sheet before starting”.

It also stops on its own: when there is no genuinely new information left, it stops asking and either writes the remaining details as assumptions or leaves them to be resolved during the conversation.

Settings

Settings → Ask Guide:

• Master switch — turn it off whenever you like;
• When to ask — same as above;
• Offer to update the task sheet when the goal changes — only reacts to explicit change phrases (“instead”, “redo”, “change it to…”), never nags every turn;
• Review the task sheet before starting;
• Max follow-up rounds;
• Brain model — the model used for the plugin’s own analysis. By default it auto-picks the newest DeepSeek model in DSH with Max reasoning; you can also choose any available model and reasoning level from the client.

Install

From the plugin market: just click install.

Manually:

dsh plugin add dsh-help-me-to-ask

Or install the tgz from a release. Restart DSH (or refresh the page) and the switch will appear next to the input box.

A few notes

• It only appears on the first message of a new conversation, and later only when an explicit change-of-mind signal appears. It never interrupts every turn.
• It does not touch your model or reasoning-level settings. I once built a version that switched models/reasoning automatically; later I realized almost nobody actually wants to pick reasoning levels for DSH, so I removed it and kept only the “ask clearly first” part.
• Subagents doing work never use this plugin.
• The analysis model is the one you choose in settings; calls go through the same route as normal conversations, and no data is uploaded anywhere else.
• Task-sheet state lives only on your own machine: $DSH_HOME/dsh-help-me-to-ask-sessions.json.
• It can be a bit clumsy sometimes. If it asks the wrong thing or a sentence feels off, tell me and I’ll keep improving it.

Changelog

See CHANGELOG.md.

Thanks

DSH’s community helped me a lot while building this, and thank you for giving it a try.

If you want to reach me, email doreihime.cn@gmail.com — I read it.

MIT License.