Transcribe image-only documents with scored multi-pass OCR and quality gating.
npx skills add https://github.com/JYeswak/grok_bot_playground --skill llm-ocr-jyeswak
Please help me install this Agent Skill. Skill: llm-ocr Source: https://github.com/JYeswak/grok_bot_playground/tree/main/plugin/skills/llm-ocr Command: npx skills add https://github.com/JYeswak/grok_bot_playground --skill llm-ocr-jyeswak
[illegible: reason] instead of guessing. - Quality gate per document: Scores every document 0-100 on legibility share, layout integrity, and character confidence, then passes, passes with warnings, or escalates with exactly what would settle it. - Clean handoff to structured extraction: Produces transcripts designed to feed doc-extract for field, table, and total logic without ever inventing values itself. - Use Case: You receive a folder of photographed receipts with no text layer. Run this Skill to get a verbatim transcript per receipt with scores, then send only the passing transcripts to structured field extraction. ## Quick Start Transcribe the scanned pages in this folder with llm-ocr and report the quality score and any illegible spans for each document.High-intent search queries and answers about installing and using this skill.
Render each page at 150 dpi minimum (300 dpi for small print), correct orientation, then transcribe visually top to bottom preserving reading order. Mark unreadable spans as [illegible] rather than guessing, and re-read flagged spans on a second pass.
llm-ocr answers what a page says and how confident the reading is, producing a scored transcript from image-only input. doc-extract answers what values are in a document, returning structured fields with provenance. Run llm-ocr first, then hand the transcript to doc-extract.
Skip OCR when the PDF already has a text layer that yields real text; say so and hand off directly to structured extraction. OCR adds no value over an intact embedded text layer and only introduces transcription risk.
Each document gets a 0-100 score from three signals: legibility share, layout integrity, and character confidence. Scores of 80 and above pass, 60-79 pass with warnings listing flagged spans, and below 60 escalates with what would settle it.
Unreadable spans are marked inline as [illegible] on the first pass and re-read on a second pass. Spans still unreadable stay flagged with a reason such as blur, glare, truncation, or handwriting; text is never invented to fill the gap.
cat-xierluoArrange images and PDF pages into multi-per-page A4 PDFs or adaptive-height long pages.
google-deepmindQuery ClinicalTrials.gov API v2 for structured trial data by condition, intervention, phase, and status.
open-circleReview API documentation routes against source code for accuracy.
xberg-ioExtract machine-readable text from scanned PDFs and images using OCR engines.