Ask

Ask

Turn unresolved decisions that affect execution into compact interactive questionnaires. Each page collects one batch of answers; handle dependencies across conversation turns rather than inside the page.

Workflow

1. Identify unresolved questions whose answers would change the approach or result.
2. Map dependencies. Put only currently answerable, mutually independent questions in the same batch. Do not implement cross-batch conditional logic in HTML.
3. Infer the language of the user's current input and select a supported locale. Honor an explicit language request when present.
4. Copy assets/example-config.json to /tmp/codex/YYYY-MM-DD/, set its permissions to 0600, then change only the batch ID, submit title, and question data.
5. Render the fixed template with scripts/render_ask.py and pass the selected locale through --locale; do not regenerate page code.
6. When a response beginning with Ask batch: arrives, update the known constraints. Generate another HTML batch only if dependent questions remain; otherwise resume the original task or briefly confirm that questioning is complete.

Question design

• Support single, multi, text, and textarea questions.
• A batch may contain multiple independent questions; each question should ask for one decision.
• Write questions and options in the same language as the selected UI locale. For unsupported languages, preserve the user's language for question content and use English for fixed UI text.
• Describe meaningful differences between options. Mark a recommendation with recommended: true, and enable free-form input with allowOther: true when useful.
• Add a preview only when it makes comparison materially clearer. Previews support html and svg for mockups or flowcharts, but export only the selected option, not internal preview state.
• Keep option layout adaptive: cards with descriptions or previews use at most two columns and collapse to one on narrow screens; short-title-only cards may auto-fit; Other spans the full row.
• Follow the active OpenAI Visualize skill's HTML contract. Inline CSS is allowed when useful; continue to exclude scripts, event attributes, iframes, network requests, and executable URLs from option previews because previews are presentational and share the questionnaire DOM.

Localization

• Built-in locales are en (English), zh-Hans (Mandarin Chinese), hi (Hindi), es (Spanish), and ar (Arabic).
• Let the AI choose the locale from the user's current input and pass it explicitly to the renderer. Do not infer language from browser settings.
• zh, zh-CN, and zh-Hans-CN normalize to zh-Hans; common regional variants of English, Hindi, Spanish, and Arabic normalize to their base locale.
• The CLI --locale value overrides an optional locale field in the config. The default is en.
• Arabic uses right-to-left layout automatically.

Rendering

In Codex:

1. Read the bundled references/visualization.md completely.
2. Write the fragment to the current thread's visualization directory:

python3 <skill-dir>/scripts/render_ask.py --config <config.json> --locale <locale> --output <thread-visualization-dir>/ask-<batch-id>.html

3. Resolve the active visualize:visualize skill directory from its available-skill source path, then validate with <visualize-skill-dir>/scripts/render.py. Do not infer the script path from the plugin root. Render with ::codex-inline-vis{file="ask-<batch-id>.html"}.

When Codex inline visualization is unavailable, generate a standalone page:

python3 <skill-dir>/scripts/render_ask.py --config <config.json> --locale <locale> --output /tmp/codex/YYYY-MM-DD/ask-<batch-id>.html --standalone

Results

• In Codex, show only the submit action. It calls window.openai.sendFollowUpMessage() and does not use the clipboard.
• In a standalone page, show only the copy-result action. It copies the Markdown answer list and does not provide submission.
• Include only the batch ID, questions, and answers in the result. Do not repeat the original task.