Prompt box

Problem Context

• Users may be asking for generation, summarization, extraction, classification, rewrite, planning, data analysis, tool use, or workflow automation.
• A prompt may include typed text, selected page content, files, current object metadata, conversation history, workspace knowledge, external tools, hidden system instructions, or no additional context.
• The AI response can be non-deterministic, incomplete, ungrounded, delayed, blocked, rate-limited, or require user review before use.
• Prompt input often sits near prompt suggestions, generated responses, chat history, citations, approvals, attachments, and model-selection controls.
• Mobile keyboards, IME composition, speech input, paste, file upload, and Enter/newline behavior can change how users submit or edit prompts.

Selection Rules

• Choose prompt box when the central user action is composing and submitting an AI request.
• Use prompt suggestions when the interface provides starter requests that should load into an editable prompt box before send.
• Use command palette when the typed input selects known app commands or routes rather than open-ended AI output.
• Use search suggestions when the input creates or refines a retrieval query and returns matching items.
• Use textarea when the field is ordinary human-authored long text without AI context, sources, model state, or generated-output lifecycle.
• Use chat interface when conversation history, assistant turns, follow-up prompts, streaming responses, and multi-turn memory are the primary surface.
• Require explicit send when a prompt can trigger tool use, publish content, contact people, alter records, reveal sensitive data, or spend money.
• Show context inclusion and repair paths before send when selected text, files, workspace sources, or tools are unavailable, stale, too large, or permission-limited.
• Keep empty-field hint copy instructional but never treat hint text as submitted prompt content.
• Preserve draft text, attachments, selected context, focus, and scroll when send fails, model availability changes, or users navigate back from output review.

Required States

• Empty prompt state with label, helpful instruction, and no implied hidden submission.
• Typing state with expandable multi-line draft, send affordance, newline behavior, and character or token boundary feedback.
• Context-attached state with selected text, current page, file, source, or tool chips.
• Unavailable-context, unsupported-file, over-limit, permission-limited, blocked-safety, and offline states with repair paths.
• Submitted prompt state that records the exact prompt sent and distinguishes it from the editable draft.
• Generation pending, failed request, rate limited, retry, edit-and-resend, cancel, and cleared draft states.
• Sensitive-content warning, external-model disclosure, tool-use review, and human-approval handoff states when relevant.
• Mobile keyboard state with accessible send, attachment, newline, voice input, and draft-preservation behavior.

Interaction Contract

• The prompt box always exposes the user's editable draft before submission.
• Send uses the visible draft and the visible context chips; hidden context must be disclosed elsewhere in the same surface.
• Placeholder, suggestion, or hint text never submits as a user prompt unless the user explicitly selects or types it.
• Enter, Shift+Enter, Command+Enter, mobile send, and IME composition behavior are stated or follow platform expectations without accidental sends.
• Context chips can be inspected, removed, repaired, or replaced before the request is sent.
• Blocked send states explain whether the problem is empty prompt, missing context, unsupported file, safety policy, permission, rate limit, model availability, or network state.
• After failed send or blocked send, the draft and attachments remain available and focus stays near the repair control.
• After submit, users can inspect the exact prompt, edit and resend it, or clear the draft without losing the previous output.

Implementation Checklist

• Define prompt payload structure: user text, selected content, files, source IDs, tool permissions, conversation scope, output format, model or mode, and safety flags.
• Keep prompt draft state separate from submitted prompt history, generated answer, command execution state, and search query state.
• Implement labelled controls for send, attach, context source, remove context, clear draft, edit previous prompt, retry, cancel generation, and open safety or privacy details.
• Validate empty prompt, over-limit text, unsupported attachments, unavailable sources, revoked permissions, offline state, rate limits, and blocked content before send.
• Preserve drafts across focus loss, route changes within the task, failed requests, file validation errors, and model availability changes.
• Test keyboard, screen reader, mobile keyboard, paste, IME composition, dictation, long prompts, file attachments, sensitive content, and low-connectivity states.

Common Generated-UI Mistakes

• Showing Ask anything as the only instruction while hiding what sources and tools the model can use.
• Auto-sending prompt suggestions, pasted content, or empty-field hint wording.
• Using a prompt box for deterministic commands that should be explicit buttons or command-palette actions.
• Styling a prompt box like search and implying generated answers are exact retrieved results.
• Letting users attach files or select context without showing whether the model can read them.
• Clearing the draft after a rate limit, upload failure, safety block, or network error.
• Hiding consequential tool use behind conversational phrasing.

Critique Questions

• Can users see the exact request, context, attachments, sources, tools, and output expectation before send?
• What prevents accidental send from Enter, starter prompts, paste, mobile keyboard actions, or empty-field hint text?
• Which failures preserve the draft, and which provide repair before submission?
• Does this surface need a chat interface, prompt suggestions, command palette, search box, or ordinary textarea instead?
• Are privacy, sensitive-data, external-model, and tool-use consequences visible at the prompt box boundary?
• Can users edit and resend the exact prompt after reviewing a weak, failed, or incomplete output?