ワンクリックで
ask-user-question
Use when missing user information materially changes the next action and the askUserQuestion tool is available.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
メニュー
Use when missing user information materially changes the next action and the askUserQuestion tool is available.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
Explicit-only Docker deploy. Run this skill's bundled scripts/deploy.sh directly; do not search for skills, inspect trees, or use subagents.
Explicit-only Docker deploy. Run this skill's bundled scripts/deploy.sh directly; do not search for skills, inspect trees, or use subagents.
Use only when explicitly invoked as $web-service-scaffold to run the bundled web-service scaffold script.
Use when a goal should be split across multiple queued session turns with reflection turns between work turns.
Enforce the repository package documentation contract as a navigable dependency graph. Use when creating or updating package README, docs pages, navigation tables, architecture router, API docs, usage docs, constraints, blast-radius/consumer records, internals, decisions, or testing docs. Docs exist so a later model drills down from task to edit target with minimal context and sees blast radius before editing. Keep content terse; keep only repo-specific contracts, paths, edges, intent, and decisions.
Enforce the repository Docker contract: Docker-first runtime, one root docker-compose.yml, module-owned Docker assets, apps service-module to container mapping, isolated networking, Docker socket passthrough for container-spawning services, a standard npm deploy entrypoint, and runtime-only Dockerfiles that copy prebuilt local dist artifacts instead of building the project. Use when defining compose services, Dockerfiles, networks, volumes, env_file wiring, deploy scripts, Docker socket access, or container ownership.
| name | ask-user-question |
| description | Use when missing user information materially changes the next action and the askUserQuestion tool is available. |
Use askUserQuestion to get information that cannot be safely inferred and
would change the next step, output, permission boundary, account choice,
configuration, or user-facing result.
Do not ask because you are merely uncertain. First inspect files, docs, logs, tool output, or durable session history when the answer can be discovered without the user.
Ask the user when any of these are true:
Do not ask when a conservative default is obvious, the missing detail has low impact, or a local inspection/tool call can resolve the question.
Ask one question by default. Use two or three only when the answers are independent and all are needed before continuing.
Every question must have:
id: stable snake_case, used only for mapping the response.header: short UI label.question: one clear question. It may include necessary context, but keep
it focused enough that the user can answer without reading a spec.inputType: single_choice, free_text, or secret.Use single_choice when the likely answers are mutually exclusive and can be
explained as choices.
Rules:
(Recommended) to the recommended option label.Use choice labels that are actionable, such as Keep current API (Recommended), Add migration, or Skip tests for now. Avoid vague labels
such as Yes, No, or Maybe unless the question itself makes the action
unambiguous.
Use free_text when the answer is an identifier, URL, policy sentence,
business rule, naming choice, or other value that cannot be represented well as
a short option list.
Ask for the smallest useful input. Do not ask the user to write a long spec when one missing value is enough.
For non-secret questions, the client may return image attachments with the answer. Use them when the user needs to show visual state, screenshots, design references, error dialogs, or UI details that would be awkward to describe in text.
If an answer contains attachments, inspect the attachment context as part of the user's answer. Do not ask for the same screenshot again unless the attachment is missing, unreadable, or contradicts the text answer.
Use secret only for credentials, tokens, passwords, or similarly sensitive
values. Never echo the value back in a final answer or tool output summary.
If a non-secret alternative exists, prefer that alternative, such as asking the user to configure an environment variable or choose an existing account.
Treat empty answers as cancellation or inability to answer. Continue only if a safe default exists; otherwise explain the blocker briefly.
When an answer includes both a selected option and a user_note: ... entry,
honor the selected option first and use the note to refine the implementation.
After receiving the answer, proceed with the task. Do not ask the same question again unless the workspace state changed or the previous answer is inconsistent with later evidence.