Tips / 🧠 Context Management

Drop the Screenshot Once the Model Has Read It

Images can be 1,000-2,000+ tokens each; removing stale ones cuts that per turn Beginner 2 min read

Multimodal chats hide a recurring cost: every image and attachment in the thread is converted to tokens and re-sent on every follow-up turn, the same way text is. A single screenshot commonly lands in the high hundreds to a couple thousand input tokens depending on its resolution, and a multi-page PDF can be far more. If you paste a screenshot, get an answer, and keep chatting, you re-pay for those pixels on turn after turn even though you're now only discussing text.

Before (wasteful):

Turn 1: [uploads a 1,600-token dashboard screenshot] "What's the error in this chart?"
Turn 2: "Okay, how do I fix that query?" (screenshot re-sent)
Turn 3: "Write a test for it." (screenshot re-sent again)

Three turns, three full charges for an image you only needed once.

After (lean):

Turn 1: [uploads screenshot] "Transcribe the error text and the failing query verbatim, then we'll continue."
Turn 2+: Continue in a fresh message (or new chat) using the transcribed text. No image re-sent.

You convert the visual into a few dozen tokens of text once, then work from the text.

Why it works: Vision models tokenize images by area or tile count, so a high-res attachment is genuinely expensive, and a stateless API re-bills the whole conversation each turn. Once the salient content is in text form, the pixels add cost without adding information. Removing them shrinks every subsequent request.

Practical moves:

  • Ask the model to extract what matters (error text, table values, layout notes) on the first turn, then drop the image.
  • In chat UIs that keep attachments inline, start a new chat once you're past the visual step and paste the extracted text.
  • Downscale before uploading when fine detail isn't needed; a 4K screenshot of a few lines of red text wastes most of its tokens.
  • Keep the image only while you're genuinely still asking about visual specifics (exact colors, spatial layout).

The rule of thumb: an image is an expensive input you should consume once and convert to cheap text, not a sticky attachment you drag through the whole conversation.

Applies to: ChatGPTClaudeGeminiOpenAI APIAnthropic API
Don't just read it — build the habit

Get a fresh tip every morning

You're reading a free Beginner tip. Pro unlocks all 38 advanced tactics and sends a new one daily — $9/mo, cancel anytime.

Unlock Pro — $9/mo More free tips

More in Context Management

🧠Context Management 50-90% on file-heavy prompts

Paste the Function, Not the Whole File

Most coding questions need 20-40 lines, not your 800-line file. Send the relevant slice plus a one-line note about the rest, and your input shrinks dramatically without hurting the answer.

Beginner Read →
🧠Context Management Trims re-sent history; often 20-60% fewer input tokens per turn after a topic switch

Start a New Chat When the Topic Changes

Chat apps re-send your whole conversation with every message. When you switch tasks, the old turns become dead weight you keep paying to re-transmit — even with caching discounts.

Beginner Read →
🧠Context Management Often 40-70% on tool-heavy agent loops

Prune Bulky Tool Results Once You've Used Them

Tool and function-call outputs are the heaviest, most disposable thing in an agent transcript. Once the model has extracted what it needs, replace the raw result with a one-line stub before the next turn.

Intermediate Read →