os-ask-simple

Two jobs. Ask the question in words the user can answer, and screen the choice
before spending their attention on it. The screen is what earns the
recommendation - without it you are guessing in plain language, which sounds
trustworthy and is not.

Language

Write in the language the user speaks in this session. Detect it from the
conversation. Keep code, file names and identifiers in English.

When to use

• You are about to ask the user a technical question.
• You are about to offer options.
• The user asks whether something is worth it, too complex, or replaceable with

something simpler.

• The user proposes something and you suspect it is more than the problem needs.

Before anything: is this even a question for them?

Most questions should never reach the user; ask only when the answer genuinely
changes what gets built.

• Can you answer it by looking? Read the code, the config, the last report.

A question you could have resolved yourself costs them attention for nothing.

• Is there a conventional default? Take it and say you took it.
• Would both answers lead to the same work? Then it is not a fork.

Light form - every question

<The question in one plain sentence. No jargon; if a term is unavoidable, give a
three-to-five word analogy.>

Why it matters: <one line, in terms of the product, not the code>
What changes later: <one line>
Easy to undo: <yes, and how - or no, and why>

Then the options, through your tool's question picker where it has one (in
Claude Code, AskUserQuestion). Offer two to four, each with a one-line
trade-off in plain words, the recommended one first and marked (Recommended).

Keep the heading to 12 characters or fewer and each label to one to five
words. Claude Code's picker has the same limits, and short labels read well as
plain text too. Where there is no picker, write the same question and options
as plain text, the recommended one first and marked.

Full form - six checks, for structural choices

Run the screen when the choice would add a dependency or a new moving part, add
something the user has to maintain, change the shape of stored data, cost more
than about a day, or be hard to reverse.

Show it as a table. Answer every row - "not checked" is allowed and honest;
silence is not.

| Check | What goes in the answer |
|---|---|
| How long now | Real effort, in hours or days, plus what has to be touched |
| Simpler substitute | The simplest thing that would also work - or "none found", having looked |
| Extra work for you later | Anything the user must do repeatedly afterwards: approvals, manual steps, watching a dashboard |
| Harder to change later | What this locks in, and what would be expensive to move afterwards |
| Over-engineering | Say yes when it is yes. A row that always answers "no" is decoration |
| Easy to undo | Reversible, and how - or one-way, and why |

Then the recommendation, in one line, as an actual opinion.

Hard rules

1. Always weigh doing nothing. "Change nothing" is a real candidate, often

the winner. If it lost, say in one line why.

1. A recommendation is required. Never lay out options and stop. "It depends"

is not a recommendation - if it truly depends, say what it depends on and pick
the option that is right under the more likely condition.

1. Recommend against the user's own idea when the screen says so. Plainly, in

one sentence, with the simpler substitute named. They asked for a filter, not
for agreement.

1. Never recommend what you have not screened. If the six checks were skipped

because the choice looked small, say the choice looked small.

1. One question at a time. Two questions in one message means the second gets

a careless answer.

1. Watch your own bias. The most interesting thing to build is not the

recommendation. If an option is more fun to implement, that is a reason for
suspicion, not for preference.

Known gotchas

• Two options that end in the same place are one option. Do not pad the

picker to look thorough.

• "Over-engineering: no" answered reflexively kills the whole screen. The row

exists to be answered yes sometimes.

• Effort estimates are guesses. Say "roughly" and give a range. A confident

number that turns out wrong costs more trust than a range ever does.

• The user may pick the option you did not recommend. That is the point of

asking. Do it their way without re-arguing, and note the trade-off once.