AI attachments: show what was sent and read

Keep the attachment visible while the question is being written.

NextGen shows a document preview directly above a typed question and keyboard, with a small removal control. AI Chat 6.0: Genius similarly keeps an image of a trading card beside an identification request. The attachment and the instruction remain one compositional unit, which makes it easier to check what “this” refers to.

The preview should survive keyboard opening, text editing and a trip to another picker. A small screen should not force the filename, remove action and send action to compete for the same touch target. For a document, show a useful identity such as filename and type. For a photo, make the preview inspectable without treating the thumbnail as a full-resolution quality check.

Removal should remove the candidate from the draft, not erase an already sent source from conversation history. Label the control with the file identity for assistive technology. If there are several attachments, allow removal of one without resetting the question or silently reordering the others. People need to be able to repair a specific mistake.

Make the sent source a durable part of the message.

Chatbot AI: AI Chat Assistant displays a document filename and size in the conversation. Pixi shows a PDF card followed by an assistant acknowledgement and follow-up suggestions. In both cases, the source has a visible place in the transcript rather than disappearing after the upload interaction.

That persistence is helpful, but an acknowledgement such as having reviewed a document is still a product statement. It is not independent evidence of extraction coverage or answer grounding. If the system knows which pages were processed, show that information accurately. If it only knows that upload completed, do not promote that fact into a claim that the full document was understood.

A sent file card should reopen the relevant source or show a clear unavailable state when access expires. Keep message identity stable when a person later attaches another version with the same filename. A history containing two reports called Final.pdf needs a way to distinguish them without depending on a timestamp hidden in a support log.

Clarify whether the person is asking a question or starting a task.

Chat AI: Ask Agent Anything puts a selected PDF, a free-form question and an optional summary choice in an Upload & Ask sheet. ChatPod instead presents a File Summary task with a selected PDF and Continue. These are related but distinct contracts: a general question about a source versus initiating a defined document operation.

When both free text and a shortcut are offered, explain how they interact. Does choosing Summary replace the typed question, supplement it or merely prefill editable text? The final request should not contain conflicting instructions that the user never intended. A reviewable prompt is often more useful than a mystery action hidden behind Continue.

In a specialist flow, carry the task name into the resulting conversation so the user understands why a response appeared. If a document is replaced, keep an explicit decision about whether the existing question still applies. An instruction about a numbered exercise should not automatically transfer to a different document just because both files are PDFs.

Explain limits before people commit sensitive or unusable input.

AI Assistant: Images & Writing names PDF and text support, an approximate ten-page limit and a sensitive-information warning before selection. Brainly shows a named image attachment with a statement that it will be used as a reference. One screen sets an input boundary; the other communicates intended use after a source is present.

Use limits your own service enforces, including any meaningful difference between file size, page count and available content. A document can be small in bytes and still exceed a page limit. Do not copy another app’s threshold into your product requirements or claim that a visible warning establishes a particular retention or training policy.

A privacy explanation should match the real processing arrangement and link to the relevant details before submission. If content leaves the device, do not imply that selecting a local file keeps analysis local. Keep warnings concise and actionable, and make a cancelled choice leave the draft intact rather than forcing the user to restart the conversation.

Keep model choice and attached input compatible.

ChatBox combines an image preview, typed instruction and model selector in one input sheet. ChatPrime similarly displays a selected PDF beside a summarization request while its model label remains visible above. These layouts expose a dependency that is easy to miss: changing the processing choice may change what inputs are supported.

Validate compatibility when the file is selected and again when the request is sent. If a user switches to an option that cannot process the source, explain the conflict beside the attachment. Do not quietly remove the file and answer the text-only question. That creates a response to a different request while preserving the appearance of success.

ScreensDesign Pro is useful here for following the full attachment sequence, not just collecting composer screenshots. Compare how a product enters the picker, returns to the draft, exposes the sent source and handles another question. Those transitions determine whether the input contract remains understandable.

Separate upload, processing and answer states.

Give each candidate attachment a stable local identity and record its name, type, byte size and upload state separately from the message. After submission, attach a server-confirmed source identity to that message. If content extraction or indexing is asynchronous, expose its state without describing an unfinished job as ready for questions.

For partial support, name the boundary: pages available, unsupported embedded media or a failed extraction. Do not manufacture a percentage when the service cannot measure progress. A retry should target the failed attachment or processing step and avoid submitting a duplicate message. Preserve the typed question when the upload fails.

Decide how source access behaves across later turns, conversation sharing and account changes. A removed draft attachment should not be treated as a deleted server copy unless that action actually performs deletion. Likewise, a historical thumbnail should not promise that the underlying source remains available forever. Describe access and retention through verified product behaviour.

Test whether the answer used the intended material.

Create controlled test documents containing easily distinguishable facts on early and late pages. Ask about a specific page, then repeat with an unsupported file, a renamed duplicate, a scanned document and an interrupted upload. Compare the visible source receipt with the material actually made available to the answering system.

Measure attachment completion, extraction failures, source-related answer corrections and retries that create duplicate messages. Separate transfer reliability from answer correctness. A high rate of successfully sent files says little about whether the user received a useful, grounded response. Evaluate both with appropriate fixtures and human review.

At large text sizes, make filenames wrap and preserve the remove control. A screen reader should announce the file, its current state and the reason Send is unavailable. After an upload error, move focus only when necessary and provide a recovery action that does not destroy the rest of the draft.

Review the attachment lifecycle.

Use these checks on a working build, not only on the polished reference state.

• ✓Keep file identity beside the draft question.
• ✓Distinguish selected, uploaded, processed and sent.
• ✓Show enforced type, page and size limits honestly.
• ✓Preserve the question on upload or processing failure.
• ✓Record source version and request compatibility.
• ✓Test partial extraction and historical access.

Related design decisions

Teach the first source-based task with the AI chat onboarding examples rather than expecting a first-time user to understand a bare attachment icon.

When source material becomes a reusable note, compare the AI note-taking app designs for the later relationship between original recording or document, generated output and saved work.