I'm transcribing The Zuni Café Cookbook (570 pages, 400+ recipes) into braille. The OCR read the two-column ingredient lists straight across the page, so every recipe came out scrambled. This workflow uses a Perplexity Computer Project to rebuild those lists in the right reading order, keep every print detail intact, and hand me a short list of spots to check. It's for anyone whose OCR has scrambled multi-column layouts.

1. Define the problem narrowly. Most of the book was fine, so the real job was: fix the reading order of two-column ingredient lists and leave everything else alone.
2. Set up a Project with standing rules. Mine cover using the attached spec as the rulebook, processing two-column ingredient lists only, keeping italics, bold, fractions, degree signs, and accented letters as printed, letting the page win when OCR and page disagree, producing only a Word doc and a review-log spreadsheet, reporting anything uncertain, and not repeating the book's text in chat.
3. Put the detail in a spec file. A Markdown file attached to the Project defines the reading order (left column top to bottom, then right), page markers like [Print page 155 | PDF page 159], file names that carry both printed and PDF page numbers, one ingredient per paragraph, and stop rules that require accounting for every page.
4. Build the spec through small pilots. Pilot 1 (15 pages) gave me the typography rules. Pilot 2 targeted the hardest layouts and showed the AI was spending effort on text my braille software already handled, so I cut the scope. I then merged everything into v1.3, with an approved page as a worked example.
5. Run each batch with the same command. I extract a page range, start a new session, attach the PDF, and paste:

"Apply Cookbook Reading-Order Repair v1.3 to the attached PDF only. Source: [FILENAME]. Printed pages [START–END]; original PDF pages [START–END]. Process two-column ingredient lists only, with their title, yield, and subheaders. Produce a complete DOCX and review-log CSV only. Do not summarize or rewrite. Report all uncertainty and any unprocessed pages. Do not repeat source text in chat."

6. Review the log, not the whole book. Each batch returns the repaired lists in Word and a CSV with one row per issue. Typical entries are a rejoined word split at a line end or a restored accent. That's 30 to 80 entries per batch instead of proofreading every page.
7. Scale up as trust builds. Batches grew from about 14 pages to more than 70. The book's two-column index had the same problem, so it went through the same process with one extra line of instruction.

It saved me from retyping most of the book by hand. The AI prepares clean text, while the braille decisions and final proofread stay with me.