PTN—07

Context Attachment

Let people ground a request in specific material — files, selections, screens — instead of describing it.

The hardest part of most prompts is not the instruction; it is the context. Describing a spreadsheet, a codebase, or a design in prose is slow and lossy, and the model fills the gaps with guesses. Context attachment replaces description with reference: drag a file in, @-mention a document, point at a selection, and the request becomes “do this to that” instead of a paragraph of scene-setting.

Attachment is also the honest way to scope an answer. A model with the actual quarterly report attached answers from the report; a model working from your summary of it answers from your summary. Products that make attachment first-class — a visible chip in the composer, not a buried upload flow — get shorter prompts, better answers, and fewer hallucination complaints, because the person controlled what the model saw.

The design work is in making attached context legible and revocable. Every attachment should render as a discrete, named object the person can inspect and remove before sending. When the system pulls context implicitly — the open file, the current page, recent history — it owes the person the same visibility. Implicit context that is invisible is how you get answers about the wrong document and a user who cannot tell why.

Anatomy

Fig. 01 — Context AttachmentLive specimen

Type @ to reference a customer, invoice, or job by name.

AUTOThis week’s schedule

1 context item will be sent with this prompt.

A composer with attached context rendered as inspectable, removable chips above the input.

1Attachment chips. Each piece of context is a named, discrete object — file name and type visible, never a mystery blob.
2Remove control. Every chip can be removed before sending. Attaching is a decision the person can reverse.
3Add affordance. An explicit attach entry point in the composer, alongside drag-and-drop and @-mention paths to the same result.
4Implicit context indicator. Context the system pulled on its own — the open file, the current selection — is shown with the same visibility as manual attachments.

When to use

The quality of the answer depends on material the model cannot otherwise see.
People are pasting long excerpts into the prompt box to give the AI something to work with.
The AI operates inside a workspace — docs, code, designs — it could reference directly.
Requests routinely start with several sentences of scene-setting before the actual ask.

Design considerations

01Render every attachment as a named chip in the composer; never absorb context silently.
02Support multiple attachment paths — drag-and-drop, a picker, @-mention — that converge on one representation.
03Let people remove any attachment before sending, including context the system added itself.
04Show implicit context (open file, current page) explicitly, and let people turn it off.
05State file-type and size limits at the point of attachment, not in an error after upload.
06Preserve attachments across a failed or edited send; re-attaching is rage-inducing.
07When an answer draws on an attachment, say which one — attachment and citation are two halves of one loop.

Pitfalls

✕Implicit context grabbing — the AI silently reads the open tab, and the person gets an answer about a document they never meant to include.
✕The mystery blob: attachments collapse into an unlabeled paperclip count, so people cannot verify what the model will see.
✕Attachment theater — the file uploads, renders a chip, and is then truncated or ignored by the model with no signal to the user.
✕Losing attachments on error or edit, forcing people to rebuild their context to retry a prompt.

In the wild

Claude: Files and images attach as chips in the composer, and Projects hold persistent context documents the model draws on across conversations.
Cursor: @-mentions pull specific files, folders, or docs into a request, so “fix this” is scoped to exact code rather than a prose description of it.
ChatGPT: Drag-and-drop file upload turns spreadsheets and PDFs into first-class conversation context, with attachments visible on the sent message.
Notion AI: Answers questions grounded in the workspace itself — pages and databases are the context, no manual upload step required.
Perplexity: Attached files scope search-backed answers to your material, blending uploaded context with cited web sources.