extracting-from-imageslisted

# Extracting From Images — Read for Intent When the user pastes a screenshot or image, your job is to **act on its content**, not narrate what's visible. "I see a screenshot showing some code" is useless. The user already knows it's a screenshot. They want you to read it. ## Iron Law **Never describe an image in the abstract. Identify what it is, extract the content verbatim, and infer why the user sent it.** If a screenshot contains code, you reproduce the code (or the relevant slice). If it contains an error, you transcribe the error. If it's a UI, you list the structure. ## Workflow 1. **Classify** — what kind of image is this? - UI / app screenshot - Code (editor / terminal output) - Error message / stack trace - Design mockup - Diagram (architecture / flow / ERD) - Chart / graph / data - Other (photo, meme, document scan) 2. **Extract** — pull the load-bearing content. OCR-style for text; structure-summary for layout. 3. **Infer intent** — why did the user share this? (Bug to fix? Implement this mockup? Explain this diagram? Replicate this UI?) 4. **Ask once if unclear** — but only if intent is genuinely ambiguous. Most of the time it's obvious from context. ## Per-Type Playbook ### Code screenshot The user pasted code as image (probably because copy-paste was awkward). - **Transcribe the code precisely.** Variable names, numbers, operators, indentation, language — all verbatim. - Then act on it (review / debug / extend / explain), depending