Topic
Running make locally is one thing. Guaranteeing that every contributor’s push or YAML edit on GitHub produces the same press-ready artifact on a clean machine with a pinned toolchain is another. How two print projects implement that guarantee with GitHub Actions.
A GUI hides complex code behind menus and drag handles you operate by hand. AI hides the same complexity behind a conversation. Recreating a hand-built InDesign leaflet as a YAML-to-LaTeX pipeline: abandoned in 2025 as beyond the time I could afford: is what made the difference concrete.
The prepare_images.py walkthrough: converting every source image to CMYK using the FOGRA39 ICC profile via Pillow and LittleCMS, with black-point compensation. Resolution is checked at placed size: never faked, never upscaled.
It’s invisible on screen. It prints correctly. But when text is extracted from the PDF — by an accessibility checker, a search engine, or a copy-paste — U+00AD surfaces as a garbage character between syllables. Here’s what it is, where it comes from, and how to find it.
Claude’s LaTeX output frequently didn’t compile. The fix wasn’t to prompt better — it was to close the feedback loop. What looked like a reliability problem turned out to be a latency problem.
Producing a PDF/X-4 press-ready file from LuaLaTeX without touching Ghostscript: the pdfx package, TrimBox/BleedBox geometry in points, FOGRA39 OutputIntent, CMYK-only color discipline, and XMP metadata — everything a print shop will verify.
One hard stop, five soft failures, a renamed file, and a report alongside it. The design philosophy behind a preflight that never aborts — plus the pikepdf/PyMuPDF six-check implementation and the --strict flag for CI.
A 370-cell calendar that shifts every year. A lead’s name in three places. An IBAN in an invisible text box. These are not unusual InDesign problems — they are what InDesign files become when they accumulate facts with no single home. Two projects show what the alternative looks like — and where InDesign’s own global tools stop short.
1,800 lecture transcripts — 1.8 million words once trapped in Word 97 files — converted to Markdown, scored and tagged by AI, and turned into a searchable discovery platform that also generates print-ready LaTeX anthologies.
