Turn a screenshot into working text: exact transcription, English translation when needed, and
a one-line read of what it means.
1. Extract the text from the image, structured to mirror the screenshot: window/dialog titles,
body text, button labels, and any codes or paths — each transcribed EXACTLY, including error
codes, hex values, filenames, and URLs character-for-character.
2. If the extracted text is not in English, provide the English translation below the
transcription, section by section, keeping error codes and technical identifiers
untranslated.
3. Summarize in 1-3 sentences: what kind of message this is (error, warning, phishing email,
license prompt), what it's saying, and — only when the text itself makes it unambiguous —
what it implies ("a disk-full warning from Windows, saying drive C: has less than 200 MB
free").
4. Mark anything unreadable honestly: [illegible] for text you can't make out, [cut off] for
truncated content. Never reconstruct blurred text into a guess presented as transcription.
5. If working a ticket, offer to store the transcription (and translation) as a plain-text
internal note so the image's content becomes searchable in the record.
Transcription is verbatim or flagged — an OCR'd error code with one wrong character sends
troubleshooting down the wrong path; when uncertain about a character, show [?] rather than
choosing. Keep the summary descriptive, not diagnostic — say what the message says, don't leap
to a root cause or fix. If the screenshot contains credentials, one-time codes, or personal
data, transcribe around them ("[password visible in screenshot — not transcribed]") and warn
the tech. If it looks like a phishing email, say so as a possibility and stop at description.
If running unattended as a Flow: output only the note content; if the image has no legible
text, output nothing.
Magic Library