AskUserQuestion Belongs in Every AI's System Prompt

The most underrated design decision in Claude Code is a small tool called AskUserQuestion. When Claude needs to clarify something, it can render an interactive question widget right in the terminal: a real, clickable list of options with descriptions, instead of yet another paragraph of prose asking what you want.

It sounds trivial. It isn't.

The problem with prose questions

Before this tool existed, the back-and-forth with an AI assistant looked like this:

AI: "To make sure I understand: do you want me to (a) refactor the auth module to use the new session API while preserving backwards compatibility, (b) refactor it and drop the legacy path, or (c) leave the existing module alone and add a parallel new one? Also, should I write tests first or just match the existing test style? And finally, would you prefer one big PR or split it into smaller ones?"

Me: reads, scrolls back, scans, types "b, match existing, smaller PRs"

That is three questions packed into a wall of text, answered with shorthand that the model now has to interpret. Half the time I forget one of the questions and only answer two. The other half I misread an option. The model fills in the gaps with assumptions, and the conversation drifts off-target before any real work has started.

What AskUserQuestion does instead

The same exchange with the widget:

Question 1: three labelled options, each with a one-line description of the tradeoff. Click one.
Question 2: two options. Click.
Question 3: two options. Click.

There is nothing to re-read and no shorthand to decode. The answers come back as a structured response that the model can act on without interpretation. You go from "translate prose into intent" to "pick a thing."

It supports multi-select for non-exclusive choices, bounded option counts so the model can't drown you in twelve checkboxes, and inline descriptions for each option's implications. You can always answer "Other" and type freely, so the widget never traps you when the right answer isn't on the list.

Why it should be in the default system prompt

Right now, models reach for AskUserQuestion inconsistently. A model will sometimes fire off a three-paragraph prose question when it could have asked three crisp multiple-choice questions. The tool is available, but the judgement of when to use it isn't reinforced.

This is fixable with a single rule in the system prompt: when the question has bounded discrete choices, use the question tool; when it's open-ended, ask in prose.

That's the whole heuristic. It's not subtle and it's not hard for a model to follow. The reason every AI assistant doesn't do this by default is, I think, mostly inertia. Chatbots have been prose-only for years, and the habit is to phrase clarifications as paragraphs.

The moment you start using bounded-choice widgets for the right questions, the conversation changes. It is faster, it is less ambiguous, and it feels more like using a tool than having a chat.

Use cases beyond clarification

The widget isn't only for "which option do you want." It works in a much wider set of situations:

Before a writing task, one widget sets tone, audience and length in one shot, with a preview of what each option sounds like.
For diagnosis, "which of these symptoms matches your bug?" narrows things down faster than "describe your bug".
Before an irreversible action, a quick "Proceed / Modify / Cancel" widget is cheaper than a long confirmation paragraph.
For a broad query, "which of these sub-topics interest you?" lets you multi-select instead of typing a list.
For prioritisation, the widget offers ranked or weighted options for "what matters most?"

All of these are bounded questions. The set of reasonable answers is small and knowable in advance. Asking them in prose wastes attention every time on a question where a click would do.

The custom instruction

If your AI doesn't reach for this tool often enough on its own, tell it to. Drop the snippet below into your custom instructions in ChatGPT or Claude.

Custom instruction

I prefer you use ask_user_input (AskUserQuestion) tool for asking me clarifications, gathering my preferences across multiple dimensions in one shot, letting me rank/prioritise options via drag-and-drop, presenting decision branches for complex tasks, doing quick gut-checks before committing to significant actions, filtering broad queries into specific sub-topics via multi-select, calibrating tone/style/audience for writing tasks, and running diagnostic/troubleshooting questions when I need help identifying or assessing something. Use it any time the question has bounded discrete choices rather than being open-ended, including simple either/or or A-vs-B follow-up questions in casual conversation. I find the interactive widget far better than answering questions in prose. For identification or self-assessment questions, lead with diagnostic questions via the tool rather than explaining criteria first.

The first line names both the OpenAI-style identifier (ask_user_input) and Anthropic's (AskUserQuestion) so the prompt works across providers. The rest is a catalogue of the situations where the widget beats prose, so the model has concrete patterns to match against rather than a vague instruction to "use the tool more."

The takeaway

The web went through this transition decades ago. Structured forms replaced free-text email confirmations because structured input is faster, cleaner, and easier to reason about for both sides. AI conversations are going through the same transition now.

OpenAI and Anthropic should put this on the default path. Until they do, the snippet above is the workaround.