Deiko
Writing Privacy Join the waitlist
Article

How to give Claude Code a screenshot

17 September 2026

Getting the image in takes four seconds and three different methods work. Then the agent changes the wrong component, and you find yourself typing four paragraphs you thought the picture had covered.

The three ways that work

Drag the file into the terminal. With Claude Code running, drag a .png from Finder or the Desktop onto the terminal window. The path lands in the prompt and Claude reads the image. This is the least fiddly method and the one that survives across terminal emulators.

Paste from the clipboard. Ctrl+V — not Cmd+V — pastes an image that is sitting on the macOS clipboard. Cmd+V is the terminal's own paste and will give you a file path or nothing at all. If you took the shot with Cmd+Ctrl+Shift+4, it is already on the clipboard.

Just say where it is. Type the path. Look at ~/Desktop/bug.png works, and it is the only method that survives being written down in a script or a task file.

Cursor and Codex differ in the keystroke and agree on the substance: the image gets in, and the image alone is rarely the thing that was missing.

Why the fix still lands in the wrong place

A screenshot answers what it looks like. It does not answer any of the questions an agent has to guess at:

  • Which element. A full-window shot of a settings page has forty controls in it. You meant one.
  • What it should have done. The picture shows the wrong state. Nothing in it says what the right state was.
  • What you did first. Bugs live in the sequence, and a still frame has no sequence.
  • What it is called in the code. You know the toggle is autoSyncEnabled. The pixels do not.

So the agent infers. Usually it infers from whatever is most visually obvious in the frame, which is why you get a confident refactor of the header when you were pointing at a checkbox below the fold.

The part that actually makes a difference

Crop to the element. A tight shot of one control removes forty wrong answers before the agent starts. It costs one extra keystroke and it is the single highest-leverage thing on this page.

Then say four things, in this order: what you pointed at, what it did, what it should have done, and where you think it lives. That is one sentence:

"The Auto-sync toggle in Settings flips back to off when I reopen the window — it should persist. Probably in the preferences store, not the view."

An agent given that plus a cropped image will usually get it right first time. An agent given a full-screen PNG and "this is broken" will not, and the three turns you spend correcting it cost more than the sentence would have.

Why nobody types it

Because it is dull, and because you already know it. The knowledge is in your head the instant you see the bug, and typing it out is pure transcription of something you have already solved. So it gets skipped, and the agent pays for it, and so do you — three turns later.

That gap is the entire reason Deiko exists. You hold a hotkey, point at the toggle, and say the sentence out loud. It crops the screenshot to what is under your cursor, reads the element's accessibility label and window title, transcribes what you said, and writes the brief. You drag one folder into Claude Code.

The screenshots and the brief never leave your Mac. The only thing that can travel is the narration audio, and only for as long as it takes to turn into text — the privacy page is specific about it.

Short version

  • Drag the file, or Ctrl+V to paste, or type the path.
  • Crop to the element. This matters more than anything else here.
  • Say what it did, what it should have done, and where it probably lives.
  • If that last step is the one you keep skipping, speaking it is faster than typing it.

Join the Deiko waitlist · More writing

© 2026 Deiko · macOS 14+
X / Twitter [email protected] Privacy Terms