detect_font_regions
ClaudeIdentify the font/typeface in an image, screenshot, photo, logo, or poster (WhatTheFont font finder / font identifier / "what font is this"). USE THIS FOR: "what font is this", "identify this font", "what typeface is used here", "which font is in this image/logo/screenshot/poster", "find this font", "name this font", "font finder", "font identifier" — with or without an image already attached. This is the ONLY entry point for font identification from an image. Call it whenever the user attaches an image and wants any font in it identified — a generic ask ("what font is this"), a named word ("identify ELEGANT"), one region, or all regions. It detects the text regions, IMMEDIATELY identifies the biggest ones (their results come back in this same call — see the
identified list — and render in the widget), and displays an interactive widget where the user browses the matches, fine-tunes regions, and identifies further regions. Pass the user's attached image in the image parameter. Also pass the user's request text verbatim in the optional user_query parameter. Do NOT use this tool for font search by description — use recommend_fonts for that. NAMED TEXT: when the user names, quotes, or clearly names ONE specific word or short phrase they want identified (e.g. "identify ELEGANT", "what font is the word Sale") — as opposed to a generic ("what font is this") or identify-all ask — ALSO pass target_text set to that exact word/phrase itself, copied from the user's OWN words, nothing else (not the full sentence). Do NOT expand it to a longer phrase or business name you see combined with it in the image — if the user says "identify cafe", pass target_text="cafe" even if the image shows "Brunch Cafe" together as one sign/logo; do not pass "Brunch Cafe". This tool matches it against the image's detected text itself and identifies that ONE region directly, in this same call. Omit target_text entirely for generic/plural/ identify-all asks. CRITICAL: - NO IMAGE YET: if the user wants a font identified but no image is available to pass (nothing attached, or the attachment is not accessible to tools), call this tool anyway with NO image arguments — pass only user_query with the user's request verbatim. It displays an upload widget where the user adds their image directly. NEVER answer with only a text request to attach or upload an image — calling this tool IS how the user gets to upload. - NEVER invent, guess, or construct a file path, URL, or id for the image — you do NOT have filesystem access to the user's attachment. This includes paths that LOOK plausible, e.g. "/mnt/user-data/uploads/...", "/home/claude/...", or any local path you did not receive verbatim from an actual tool result. If you don't already have a real image reference from a previous result, treat it as NO IMAGE YET (above) instead of fabricating one. - ALWAYS run this tool FIRST for a newly attached image the user wants identified. If it has not run for that image, call it yourself with that image; do NOT ask the user to use the widget or re-upload. You need its result (the image reference — image_id or file_id — plus the regions list) before any predict_font call. - MULTIPLE IMAGES: this tool works on ONE image at a time. If the user attached more than one image, do NOT ask which to start with — immediately call this tool for the FIRST image in the same turn (unless the user's message clearly names a different one — then start with that one). In your response, briefly note that identification runs one image at a time, you are starting with that image, and you'll continue with the next when this one is done. Complete each image's identification before starting the next; never call this tool for two images in the same turn. - NEVER announce that analysis is starting, say "please wait", or describe what will happen without actually emitting the tool call in the same turn — a response about analyzing an image is invalid unless it contains the detect_font_regions call. - If you cannot access a previously uploaded image when continuing to a later image, do not claim t…





