doc-eyelisted

Audits a Word (.docx), PowerPoint (.pptx), or PDF document the way a human reviewer would — renders every page as an image and runs a vision-capable Claude over each page to catch formatting drift, typography mixing, broken cross-references, SmartArt overflow, blank pages, colour rot, orphans/widows, placeholder text, and visual incoherence. Use after generating a document, before sending it, or when "the code is right but the rendered output looks wrong". Triggers — "audit this PDF", "audita este Word", "review the rendered doc", "doc-eye", "check the document visually", "look at the rendered output", "the PDF looks off", "is the layout broken", "está bem formatado", "verifica o documento". NOT a content reviewer for clinical correctness or scientific argument (use scientific-manuscript-reviewer) — this is purely about visual + structural fidelity of the rendered page.
pedroreisper/doc-eye · ★ 0 · Data & Documents · score 73

Install: claude install-skill pedroreisper/doc-eye

# doc-eye — vision-LLM audit of rendered document pages Claude is excellent at code-level edits to source files but blind to the visually-rendered output. SmartArt clips at the page edge. Fonts drift across sections. A heading falls into a widow on page 7. The XML looks correct; the PDF looks wrong. This skill closes that gap. It takes a `.docx`, `.pptx`, or `.pdf`, renders every page to a PNG at print DPI, and feeds each page to a vision-capable Claude with a structured rubric. The output is a JSON report (machine-readable) plus a prose render (human-readable) that lists every finding with page number, location, severity, and a concrete description. The skill reports — it does not modify the source file. It is a reviewer, not a fixer. You decide what to do with the findings. ## Core pipeline ``` ingest → render → per-page vision audit → aggregate + dedup → report ``` 1. **Ingest** — detect format (`.docx`/`.pptx`/`.pdf`), reject encrypted PDFs early, copy to a temp workspace. 2. **Render** — DOCX/PPTX → PDF via `soffice --headless`, then PDF → PNG per page via PyMuPDF (200 DPI default). Output: `pages/page-001.png … page-NNN.png`. 3. **Per-page vision audit** — one Anthropic API call per page to Claude (Haiku 4.5 by default; Opus 4.7 with `--opus`). Forced tool use returns structured JSON per page. 4. **Aggregate + dedup** — collapse repeated findings (e.g. same broken header on every page → one document-level finding with `pages: [1,2,3,…]`). 5. **Report** — write `au