keryx cards screen¶
Description¶
OCR-screen a workspace's card illustration takes for accidental text (R-GEN-26).
Image models sometimes render stray letters or words into the art, which look wrong
in a promo reel. cards screen sends each take to the image provider's vision
capability with an OCR prompt and flags takes with detected glyphs.
It's the automatic pass; keryx cards sheet (the contact-sheet PNG) stays the human
visual fallback. Flags feed the reel-make auto-pick, which prefers a clean take
(reel make runs this screen automatically after generating card takes).
- No system dependency — screening uses the configured image provider (default
Gemini), like every other backend. With no vision provider configured the takes
are reported unscreened (fall back to
cards sheet). - No re-spend — verdicts are cached per take by content hash (in
cards/takes/screen.json, git-ignored + pruned), so re-screening an unchanged take never re-calls the provider.--forcere-OCRs.
Usage¶
--workspace, -w— the reel workspace.--card N— screen one card's takes (default0= every card).--force— re-OCR even cached takes.
Example:
$ keryx cards screen -w my-post
card 3 take 1 clean
card 3 take 2 ⚠ text: CODE
card 5 take 1 clean
1 of 4 take(s) flagged
Run keryx cards screen --help for the authoritative, always-current flag set.