PDF

pdf skill

Purpose: Give the agent a small, deterministic set of CLIs for the common PDF operations: render Markdown to a well-typeset PDF, merge PDFs, split them by page range or into individual pages, and (via references) extract text or fill forms. Picking the right library on the fly is the single biggest source of PDF bugs; delegating to scripts that embed those choices removes the variance.

1. Red Flags (Anti-Rationalization)

STOP and READ THIS if you are thinking:

• "I'll use pypdf to extract the text." → WRONG for layout-dependent content. pypdf's text extraction is famously unreliable on anything with columns or complex layout; use pdfplumber. See references/library-selection.md.
• "I'll improvise a pdfplumber script to convert this PDF to Markdown." → WRONG to improvise from scratch. Run pdf_extract.py for the structured dump and follow references/pdf-to-markdown.md — and never skip the scan check: on a scanned PDF pdfplumber returns empty text silently; pdf_extract.py exits 10 instead.
• "I'll reach for playwright for Markdown → PDF because it handles everything." → WRONG. Playwright pulls a 200 MB Chromium install. weasyprint handles 95% of Markdown/HTML inputs with a fraction of the footprint.
• "I'll fill this XFA form with pypdf." → WRONG. pypdf doesn't fill XFA — only AcroForm. Detect the form type first (see references/forms.md) and fail loudly if it's XFA rather than silently writing an unchanged file.
• "I'll skip checking exit codes from pdf_merge.py." → WRONG. Missing an input file produces exit 1 and the output is absent — silently assuming success ships a broken deliverable.
• "This HTML carries its own @media print and @page rules, so html2pdf.py will honour them." → PARTLY WRONG, and the two engines fail differently. weasyprint does apply @media print and @page (measured: a 41-slide deck rendered 41 pages at its declared 297×176 mm) but supports no container queries — a layout whose type scale is built on cqw collapses into overlapping text on most pages. --engine chrome has the layout engine but forces media: screen and scales to fit the paper (§2), so an authored print stylesheet never applies at all. For a document authored for print — its own @page size, its own break rules, modern layout — drive headless Chrome directly (--headless=new --print-to-pdf), which keeps media: print, then verify with preview.py.
• "The PDF has the right page count, so the render is fine." → WRONG. Page count catches split pages and nothing else. Render every page with preview.py and look: web fonts silently fall back to system faces, and a @media (max-width:…) block with no screen qualifier also applies to print, collapsing multi-column layouts.
• "I'll grep the PDF bytes for /BaseFont to confirm the fonts embedded." → WRONG. Font descriptors live in compressed object streams, so the grep reports only the few uncompressed entries and shows system fallbacks even when the real fonts are embedded. This misreads a working file as broken. Judge embedding from a rendered page, or with pypdf.

2. Capabilities

• Render Markdown (+ optional custom CSS) to a typeset PDF via weasyprint. Fenced ```mermaid blocks pre-render to PNG via mmdc; bundled scripts/mermaid-config.json ships an office-friendly Cyrillic-capable font stack (override with --mermaid-config PATH, opt out with --no-mermaid-config). Inline $…$ and display $$…$$ **math** pre-render to MathML via the bundled katex (weasyprint typesets MathML natively; it runs no JS, so client-side KaTeX/MathJax can't be used) — currency ($5) and $ inside code/fences are left untouched, and without node/KaTeX formulas degrade to literal text (opt out with --no-math, fail-hard with --strict-math). The PDF carries a navigable **outline (bookmarks)** auto-built from h1–h6 headings — no flag needed.
• Render HTML / web archives to PDF via html2pdf.py — same weasyprint pipeline, natively handles .html/.htm, .mhtml/.mht, and .webarchive. Validated across Fern (OpenRouter), Mintlify (Anthropic Claude Code, Discord, Berachain), GitBook (Hyperliquid), Confluence (Atlassian wikis), Хабр, vc.ru and generic blogs — 34/34 fixtures pass in both modes. Bundled stylesheet on by default; --no-default-css for fully-styled inputs (BI dashboards, branded reports); --css EXTRA.css stacks on top. --reader-mode extracts the main article body (Safari Reader View parity). --timeout 180 SIGALRM watchdog with $HTML2PDF_TIMEOUT override. Universal preprocessing handles draw.io/Confluence SVG diagrams, table-based code blocks (Fern/Mintlify shiki), Tailwind/FontAwesome icon strip, ARIA-role tables (GitBook), ad-network removal, and pathological-CSS protection (Хабр content-drop bug, vc.ru CPU-loop bug). Full pipeline + flag semantics + per-platform notes documented in references/html-conversion.md. Output PDFs carry a navigable outline (bookmarks) from h1–h6 headings — engine-agnostic (weasyprint and --engine chrome; the chrome engine emits a tagged PDF, the mechanism Chromium uses for the outline).
• Merge multiple PDFs into one preserving bookmarks (pdf_merge.py).
• Split a PDF by explicit page ranges, one-per-page, or fixed-size chunks (pdf_split.py).
• Stamp a text or image watermark on every (or selected) page via pdf_watermark.py (drafts, "CONFIDENTIAL", brand stamps). --position center|top-left|top-right|bottom-left|bottom-right|diagonal, --opacity, --rotation, --pages "1-5,8". Builds one overlay per unique page mediabox, so heterogeneous decks (Letter+A4) keep correct proportions.
• Detect, inspect, and fill AcroForm fields via pdf_fill_form.py — three modes: --check (form-type triage with exit codes 0/11/12 = AcroForm/XFA/none; 13 = the stdout reader went away), --extract-fields (dump field schema as JSON for editing), and fill mode (INPUT.pdf DATA.json -o OUT.pdf [--flatten]). XFA forms are detected and refused with a clear message.
• Extract text, tables, and layout via pdfplumber (documented; inline usage from the agent is fine).
• Dump a PDF's per-page text + tables to structured JSON via pdf_extract.py — a structured dump, NOT a Markdown converter (it never emits Markdown). Its defining feature is scan detection: an image-only document exits 10 with a DocumentScanned signal instead of silently yielding empty text. Robust word-splitting by default: LaTeX/academic two-column PDFs encode inter-word spacing as positional gaps (no space glyphs), which pdfplumber's absolute tolerance glues into ASurveyonBlockchain; a font-relative x_tolerance_ratio (default 0.15) splits them correctly without regressing real-space PDFs (tune/disable via --x-tolerance-ratio R, see references/pdf-to-markdown.md §3.8). Two further silent-loss signals at exit 0: figure_pages — a page that is mostly diagram with too little text to be a text page, which the absolute char threshold misses as soon as the page carries a running header (§3.4); and text_layer_lossy — the document embeds no fonts and every encoding is single-byte Latin, so any non-Latin text was destroyed when the file was written and OCR cannot recover it (§3.9). Two grouping knobs for pdfplumber defaults that misread real documents: --y-tolerance PT (a list marker in a smaller point size otherwise becomes its own line, sorted after its item — §3.1) and --table-strategy lines_strict (background shading otherwise becomes a phantom table, or a bogus extra row on a real one — §3.2). Hyperlinks are content, and they used to vanish: every page now carries links ({uri, text, bbox} per /URI annotation, with the text the link covers) and the top level a link_count — measured at 578 links across a 20-document corpus that the dump dropped entirely, which for a web page printed to PDF is losing content, not formatting (§5). Two further advisory counters in layout_hints: split_table_rows (a table row cut in half by a page break arrives with an empty first cell while its label stays behind on the previous page — §3.3) and multi_column_pages (a full-height column gutter, whose x coordinate the hint reports because the repair is a crop, not a flag: --layout does not separate columns — §3.1). Artwork extraction is opt-in via --extract-images DIR: embedded rasters are copied out byte-for-byte and vector figures (diagrams with no image object) are cropped from the page through Poppler, each listed per page as images so Markdown references real paths — the second half of the figure_pages repair (references/pdf-to-markdown.md §3.10). Pairs with references/pdf-to-markdown.md for the PDF→Markdown decision tree and recipe; final Markdown composition stays agent judgement.
• OCR a scanned (image-only) PDF into a searchable PDF via pdf_ocr.py — wraps ocrmypdf to overlay an invisible OCR text layer (default languages eng+rus), the remediation hop for pdf_extract.py exit 10. The OCR engine is soft-optional: install with bash scripts/install.sh --with-ocr (+ system tesseract/eng/rus/ghostscript); a missing engine or language pack fails loud, never silent. See references/ocr.md.
• Render any .pdf (or peer-skill .docx/.xlsx/.pptx) into a single PNG-grid preview via preview.py (uses Poppler directly for .pdf; LibreOffice + Poppler for OOXML).
• Emit failures as machine-readable JSON to stderr with --json-errors (uniform across all four office skills).

3. Execution Mode

• Mode: script-first for the bundled operations, prompt-first with library references for extraction and form filling.
• Why this mode: The bundled operations (render, merge, split) are stable recipes. Extraction and form filling depend heavily on the specific document and deserve inspection before running — the references guide the inline work.
• pdf_extract.py — the bounded exception: extracting per-page text + tables to a JSON dump IS a stable recipe, so it is bundled (it also closes the silent-scan failure with code). Markdown composition — heading levels, reading order, table stitching — stays prompt-first agent judgement: there is no Markdown-converter script, by design. See references/pdf-to-markdown.md.

4. Script Contract

• Commands:
  • python3 scripts/md2pdf.py INPUT.md OUTPUT.pdf [--page-size letter|a4|legal] [--css EXTRA.css] [--base-url DIR] [--no-mermaid] [--strict-mermaid] [--mermaid-config PATH | --no-mermaid-config] [--no-math] [--strict-math]
  • python3 scripts/html2pdf.py INPUT OUTPUT.pdf [--page-size letter|a4|legal] [--css EXTRA.css] [--base-url DIR] [--no-default-css] [--reader-mode] [--archive-frame N|main|all|auto] [--list-frames] [--timeout SECONDS] [--engine weasyprint|chrome] [--chrome-js] — INPUT may be .html/.htm, .mhtml/.mht, or .webarchive; sub-resources in archives are extracted to a temp dir automatically. --reader-mode extracts the main article content (Confluence-priority candidate list with body-ratio guard for <main>; longest-match per selector handles archive pages with multiple .entry divs and Disqus comment threads + title-match LCS bonus for multi-article feed pages), stripping navigation, ads, sidebars, and SPA chrome (ARIA role=navigation\|complementary\|banner\|contentinfo + semantic <aside>/<nav>/<footer> + shallow <header>) — ideal for browser-saved news/blog/docs pages and hydrated SPAs. --archive-frame N|main|all|auto (pdf-8, 2026-05-05): selects which inner frame in webarchive/MHTML to render — main (default) = main resource only; N (1-indexed) = specific inner frame; all = concat all "substantial" frames (≥ 1 KB + 0 <script> + ≥ 30 chars text + not single-<img>-only) with <hr><h2>Frame N</h2> separators + per-frame namespace + sha1 image-dedup + encoding parity; auto = deterministic (0 substantial → main, 1 → that frame with main-dominance guard, 2+ → all). Vendor-agnostic: validated on 9 real fixtures across Angular/Closure/Framer/bare-DOM SPA stacks without a single vendor name in the heuristic. --list-frames (pdf-8): prints inner-frame inventory (index/kind/substantial/bytes/scripts/text-len/url) and exits without rendering — for picking N deterministically. --timeout (default 180s, $HTML2PDF_TIMEOUT env, 0 disables) caps weasyprint render via signal.SIGALRM for pathological inputs. Exit 1 with RenderTimeout envelope on watchdog fire; exit 2 with NoSubstantialFrames / FrameIndexOutOfRange envelopes for archive-frame errors. --engine weasyprint|chrome (pdf-11, 2026-05-05): render engine selector. weasyprint (default) — pure-Python typeset PDF, no browser runtime. chrome — opt-in headless Chromium via Playwright (150 MB, install with bash install.sh --with-chrome). Use chrome when weasyprint produces broken output: Material 3 calc/var bugs (Gmail-class), Framer infinite layout loops, ELMA365 inline.py assertion, JS-hydrated content, <canvas> charts. Chrome path skips weasyprint preprocess (calc-strip, font-face-strip, NORMALIZE_CSS) — those are weasyprint workarounds Chrome doesn't need; reader-mode and --css EXTRA.css remain engine-agnostic. Chrome render hardenings (universal layout strategy, post-VDD-iter-8 — 8 adversarial iterations): <script> strip from HTML + JS-enabled at context level (page can't run own JS → no Gmail self-destruct, no Angular half-hydration; we keep page.evaluate for surgical DOM normalization; --chrome-js opts page-JS back in for canvas/hydration); <base href> stripped (webarchives carry <base href="https://orig-site/"> which would route every relative URL to the offline-blocked origin); media forced to screen (default print triggers nav-hiding @media print rules in SPAs); 1280×1024 viewport (desktop CSS); layout-normalize CSS — high-specificity body release, icon-font ligature suppression with :not(:has(*)) leaf-only guard (avoids font-size:0 cascading through CSS inheritance to children), `[class="spinner"]exact-word match (not substring — avoids hidingclass="spinner-class-banner"and similar), image cap (200px), avatar-image cap (48×48 only on... img, not bare class so containers aren't shrunk); JS-based DOM normalize via page.evaluate— width-gateoffsetWidth ≥ 200 for overflow release (narrow icon-sidebars at 64px stay clipped, no label leak), substantial-modal release (position:fixed → staticonly when wide AND tall AND text-rich), modal-portal hide (when modal released, hide non-portal body children to remove underlying CRM page); **scale-to-fitpage.pdf(scale = pdf_usable / viewport_width)≈ 0.561** so 1280 px layout fits A4's ~718 px usable width without right-edge cutoff. Validated on 3 SPA archive shapes × 2 modes (Gmail Closure email, ELMA365 Angular dashboard, Yandex Cloud Console marketplace) — all 6 combinations produce full content with no overlap, no cutoff, no chrome-icon ligatures, no underlying-page noise. **Recommended composition**:--engine chrome --reader-modefor email/newsletter/article archives (cleaner article-only render);--engine chromealone for dashboards/registries/structured UIs (preserves card layout). Exit 1 withChromeEngineUnavailable` envelope if Playwright not installed.
  • python3 scripts/pdf_merge.py OUTPUT.pdf INPUT1.pdf INPUT2.pdf [INPUT3.pdf ...]
  • python3 scripts/pdf_split.py INPUT.pdf --ranges "1-5:part1.pdf,6-10:part2.pdf"
  • python3 scripts/pdf_split.py INPUT.pdf --each-page OUTDIR/
  • python3 scripts/pdf_split.py INPUT.pdf --every N OUTDIR/
  • python3 scripts/pdf_watermark.py INPUT.pdf OUTPUT.pdf (--text "DRAFT" | --image STAMP.png) [--opacity 0.3] [--position center|top-left|top-right|bottom-left|bottom-right|diagonal] [--rotation 45] [--font-size 60] [--color "#888"] [--scale 0.5] [--pages "all"|"1-5,8,12-end"]
  • python3 scripts/pdf_fill_form.py --check INPUT.pdf — exit 0/11/12 = AcroForm/XFA/none, 13 = stdout went away. (Custom codes start at 10 to leave 0–9 for argparse / shell convention. Every JSON payload this script puts on stdout goes out as UTF-8 bytes, so it does not change with the caller's locale — --extract-fields -o FILE writes its JSON to the file instead, which was already UTF-8. A reader that goes away — | head, or a consumer that exits first — ends any of those writes as exit 13 + an OutputWriteFailed envelope, never a traceback and never the interpreter's substituted 120; that covers the one human-readable status line --extract-fields -o FILE prints too. Not covered on the locale axis: that status line is presentation, not JSON, so it is still encoded with the caller's codec — a character the codec cannot render is escaped, not fatal.)
  • python3 scripts/pdf_fill_form.py --extract-fields INPUT.pdf -o fields.json
  • python3 scripts/pdf_fill_form.py INPUT.pdf DATA.json -o OUTPUT.pdf [--flatten]
  • python3 scripts/preview.py INPUT OUTPUT.jpg [--cols 3] [--dpi 110] [--gap 12] [--padding 24] [--label-font-size 14] [--soffice-timeout 240] [--pdftoppm-timeout 60]
  • python3 scripts/pdf_extract.py INPUT.pdf [-o OUT.json] [--layout] [--password PW] [--x-tolerance-ratio R] [--y-tolerance PT] [--table-strategy lines|lines_strict] [--extract-images DIR] [--image-dpi N] [--no-vector-images] [--json-errors] — dumps per-page text + tables as structured JSON (NOT Markdown). --x-tolerance-ratio (default 0.15) is the font-relative word-split threshold that un-glues LaTeX/academic PDFs; 0 disables it (legacy absolute tolerance). --y-tolerance (default: pdfplumber's 3 pt) is the absolute line-grouping threshold — raise it to 5 when a smaller-point-size list marker is split onto its own line and sorted after its item. --table-strategy (default lines) selects the table-edge strategy; lines_strict ignores fill-only background rectangles, which otherwise fabricate tables out of shading. The dump names its input in top-level source (resolved, so a relocated dump still says which PDF it describes), echoes all three effective values plus a document-level fonts list, and carries two further exit-0 loss signals — figure_pages (page is mostly artwork: per-page image_coverage / vector_coverage ≥ 0.25 with < 200 chars) and text_layer_lossy (no embedded font, no /ToUnicode, all-Latin encodings → non-Latin text was lost at export; OCR does not help). Both warn on stderr and leave the exit code alone. Every page also carries links — one {uri, text, bbox} record per /URI annotation, text being the characters the link covers (null over an image; internal /GoTo links are deliberately not reported) — and the top level a link_count. layout_hints (always present) counts four things: orphan_list_markers (a bullet in a smaller point size becomes its own line and sorts after its item), single_column_tables (background shading read as a table), split_table_rows + split_table_pages (a row severed by a page break: its continuation opens the next page's table with an empty first cell, and the stranded label — 2.11 — is named where the previous page's text still holds it; the dump does NOT stitch, that is composition) and multi_column_pages + column_probe (a full-height gutter, reported with its x coordinate and with what cropping there measured, because --layout preserves the interleaving rather than undoing it). While the matching knob is still at its default, the script tries that knob on up to three affected pages and reports what it measured — --y-tolerance 5 reunited N of M, or changed nothing (an arXiv export interposes a line between marker and item; a Confluence export's gap is wider), --table-strategy lines_strict removed N of M, or kept them all (a real one-column table, a ruled box, or a fragment of a split wider table). Advisory only; the exit code never moves, the counts stay in the dump when the line is suppressed, and the probe costs ≤0.12 s and does not run when no hint would fire. --extract-images DIR writes the document's artwork out and lists it per page as images ({file, kind, bbox, name, width, height, bytes, sha1}) — two classes: embedded rasters copied byte-for-byte, and vector figures (diagrams drawn with path operators, for which no image object exists) cropped from the page via Poppler at --image-dpi (default 150). Classification is by object model, never by looks: a block diagram that looks vector is often an RGBA PNG, and the raster branch covers it. Identical images are written once (sha1 dedup — one measured document placed a single backdrop 31 times), page-sized rasters are skipped (a background wash, or a scan whose repair is OCR), a raster declaring more than 80 MP is refused before decoding (pypdf allocates from the declared /Widthx/Height), and three further rejections keep the directory usable: an inline glyph (an emoji from a colour font is a raster XObject — square, as tall as its text line and adjacent to it; 10 of 15 files on one measured document), a flat-colour raster (a blank page's only image, which also puts that page in top-level blank_pages so the scan warning can say there is nothing to OCR) and a vector cluster enclosing the page's body text (table ruling whose crop is a picture of the page — 345-385 KB of it on four measured pages; preview.py is the fallback). DIR is mandatory, and what did NOT come out is reported in images_summary + stderr — including the case where nothing came out, which is the normal outcome for an OCR'd scan (every page is one page-sized raster, the backdrop rule drops it): the run says so instead of leaving an empty directory and a silent stderr. Every bbox — raster and vector alike — is in page coordinates. The one silent omission: a vector figure drawn entirely with fills and no stroked path is not extracted and its page is not figure_dominant either (a flat pie measures 0.07 against the 0.25 threshold) — render the sheet with preview.py for those. Without the flag the dump is unchanged (no images key at all). Exit codes: 0 success; 1 failure (missing / not-a-PDF / corrupt / encrypted-without-password / --extract-images names a file); 2 usage error; 6 SelfOverwriteRefused (-o or --extract-images resolves to the input PDF); 10 DocumentScanned — the whole document is image-only, run OCR or read the pages as images. On exit 10 the dump is still emitted; exit 10 + stderr is the loud signal. Default output is stdout; -o writes a file (idempotent). The stdout dump is UTF-8 whatever the locale says, and a pipe closed early (| head) exits with the code its envelope declares. See references/pdf-to-markdown.md.
  • python3 scripts/pdf_ocr.py INPUT.pdf OUTPUT.pdf [--lang eng+rus] [--skip-text|--redo-ocr|--force-ocr] [--sidecar OUT.txt] [--jobs N] [--password PW] [--deskew] [--rotate-pages] [--clean] [--json-errors] — OCR a scanned PDF into a searchable PDF via ocrmypdf (default languages eng+rus). --password decrypts an encrypted input; --rotate-pages needs tesseract osd data; --clean needs unpaper. Exit codes: 0 success; 1 failure (type in the envelope: OcrEngineUnavailable / LanguagePackMissing / EncryptedInput / InputUnreadable / PriorOcrFound / OutputWriteFailed / InputNotFound); 2 usage; 6 SelfOverwriteRefused. Soft-optional engine — bash scripts/install.sh --with-ocr first. See references/ocr.md.
  • All scripts above accept --json-errors to emit failures as a single line of JSON on stderr ({v, error, code, type?, details?}). The schema version v is currently 1; argparse usage errors are routed through the same envelope (type:"UsageError").
• Inputs: positional paths; optional flags per command.
• Outputs: single PDF files (md2pdf, pdf_merge) or multiple PDFs under a directory (pdf_split). All stdout goes to the output path list.
• Failure semantics: non-zero exit on missing inputs, invalid range specs, or library errors. Error detail to stderr.
• Idempotency: all three scripts overwrite their outputs on re-run.
• Dry-run support: not applicable.

5. Safety Boundaries

• Allowed scope: only paths named on the command line.
• Default exclusions: do not fetch remote resources unless the user explicitly provides URLs; md2pdf.py --base-url defaults to the input's directory.
• Destructive actions: all three scripts overwrite their outputs without prompting.
• Optional artifacts: custom CSS via md2pdf.py --css is optional; defaults produce a reasonable layout.

6. Validation Evidence

• Local verification:
  • python3 -m venv .venv && source .venv/bin/activate && pip install -r scripts/requirements.txt — installs pypdf, pdfplumber, weasyprint, markdown2, reportlab.
  • bash scripts/tests/test_e2e.sh — runs the end-to-end smoke suite (md2pdf, merge, split, fill-form, mermaid, pdf_extract, pdf_ocr). Includes the html2pdf regression battery: ~37 unit tests for html2pdf_lib/ helpers + data-driven fixture battery (6 synthetic micro-fixtures + 6 hand-stripped real-platform slices + N tmp/ originals when present on disk; per-fixture page-count / size / required+forbidden-needle assertions, see tests/battery_signatures.json).
  • Adding a new platform fixture (e.g. you found a Notion/Stripe page that breaks): drop the .webarchive/.html/.mhtml file into tmp/, run python3 scripts/tests/capture_signatures.py (auto-captures page count + needles + size band; only ADDs new fixtures unless --refresh is passed), hand-add chrome strings to forbidden_needles in battery_signatures.json, commit the JSON delta. Total ~5 min per new site. Detailed in references/html-conversion.md §Regression coverage.
  • python3 scripts/md2pdf.py examples/fixture.md /tmp/invoice.pdf --page-size letter — produces a non-empty PDF.
  • python3 -c "from pypdf import PdfReader; r=PdfReader('/tmp/invoice.pdf'); print(len(r.pages))" — returns at least 1.
  • python3 scripts/pdf_merge.py /tmp/merged.pdf /tmp/invoice.pdf /tmp/invoice.pdf && python3 -c "from pypdf import PdfReader; print(len(PdfReader('/tmp/merged.pdf').pages))" — 2× the page count.
  • python3 scripts/pdf_split.py /tmp/invoice.pdf --each-page /tmp/pages/ — produces /tmp/pages/invoice-001.pdf.
• Expected evidence: /tmp/invoice.pdf, /tmp/merged.pdf, /tmp/pages/invoice-001.pdf.
• CI signal: python3 ../../.claude/skills/skill-creator/scripts/validate_skill.py skills/pdf — exit 0.

7. Instructions

7.1 Pick the library, not the script first

A full PDF→Markdown converter is deliberately not bundled — Markdown composition (heading levels, reading order, stitching a table across pages) is agent judgement. Form filling likewise depends on the document.

1. Check references/library-selection.md for which library matches the task.
2. For PDF → Markdown: follow references/pdf-to-markdown.md — its decision tree picks digital-vs-scanned, and pdf_extract.py gives a structured dump. You compose the Markdown from that dump; the script never emits Markdown.
3. For other extraction (a one-off text/table grab): write inline pdfplumber code, or run pdf_extract.py for a quick structured dump.
4. For form filling: follow references/forms.md — detect AcroForm vs XFA first.

7.2 Creating PDFs from Markdown

1. python3 scripts/md2pdf.py input.md output.pdf covers the common case.
2. Pass --css custom.css when the user provides brand styling.
3. For images referenced with relative paths, either put them next to the Markdown file or pass --base-url /absolute/image/root.
4. For HTML-heavy inputs (embedded <style>, flexbox, columns), weasyprint handles those in the script — no extra work needed.

7.3 Merging PDFs

1. Order matters: python3 scripts/pdf_merge.py out.pdf file1.pdf file2.pdf file3.pdf appends in that order.
2. Bookmarks from each input are preserved and nested under a parent named after the source's stem.

7.4 Splitting PDFs

Three modes, exclusive:

• --ranges "1-3:intro.pdf,4-8:body.pdf,9-12:appendix.pdf"
• --each-page OUTDIR/ — one PDF per input page, zero-padded filenames.
• --every N OUTDIR/ — chunks of N pages each.

Page numbers are 1-indexed and inclusive. Invalid ranges exit 1.

7.5 Setup

1. MUST run bash scripts/install.sh once. It creates scripts/.venv/ locally, installs requirements.txt, probes whether weasyprint can find its native libraries, and prints install hints if not. Idempotent.
2. External system libraries (checked by install.sh, installed manually per project plan §3.3 "внешние инструменты — не бандлятся"):
   • pango, cairo, gdk-pixbuf — weasyprint native runtime; required by md2pdf.py. macOS: brew install pango gdk-pixbuf libffi. Debian: sudo apt install libpango-1.0-0 libpangoft2-1.0-0 libharfbuzz0b libcairo2 libgdk-pixbuf2.0-0. See references/weasyprint-setup.md for fuller notes.
   • tesseract (+ eng/rus data) and ghostscript — only for pdf_ocr.py; installed by bash scripts/install.sh --with-ocr (which installs ocrmypdf into the venv and probes these). macOS: brew install tesseract tesseract-lang ghostscript. Debian: sudo apt install tesseract-ocr tesseract-ocr-eng tesseract-ocr-rus ghostscript. See references/ocr.md. Commands that need them fail with a clear error until installed.

8. Workflows (Optional)

Markdown-driven PDF:

- [ ] Draft the Markdown content
- [ ] `python3 scripts/md2pdf.py doc.md doc.pdf`
- [ ] Open the PDF, check layout (orphans/widows, table page breaks)
- [ ] Iterate on CSS if needed (`--css brand.css`)

Merge + split for distribution:

- [ ] `python3 scripts/pdf_merge.py combined.pdf intro.pdf body.pdf appendix.pdf`
- [ ] `python3 scripts/pdf_split.py combined.pdf --each-page out/`  (if per-page delivery is needed)
- [ ] Verify page count with pypdf or Preview

Extract text (inline, no bundled script):

- [ ] Read references/library-selection.md, pick pdfplumber
- [ ] Inline: open the file, call page.extract_text(layout=True)
- [ ] For tables, page.extract_tables() with appropriate snap_tolerance

9. Best Practices & Anti-Patterns

Rationalization Table

10. Quick Reference

11. Examples (Few-Shot)

Fixture: examples/fixture.md.

Input — user request:

Turn this invoice Markdown into a letter-sized PDF.

Output — agent action:

python3 scripts/md2pdf.py invoice.md invoice.pdf --page-size letter

Input — user request:

Join these three quarterly reports into one annual PDF.

Output — agent action:

python3 scripts/pdf_merge.py annual.pdf q1.pdf q2.pdf q3.pdf q4.pdf

Input — user request:

Split the 120-page handbook into chapters of roughly 10 pages each.

Output — agent action:

python3 scripts/pdf_split.py handbook.pdf --every 10 chapters/

Input — user request:

Convert this Confluence-export HTML report to PDF.

Output — agent action:

python3 scripts/html2pdf.py q1-report.html q1-report.pdf --base-url ./q1-report_files/

Input — user request:

Stamp every page of this draft contract with "DRAFT" diagonally.

Output — agent action:

python3 scripts/pdf_watermark.py contract.pdf contract-draft.pdf --text "DRAFT"

12. Resources

• references/library-selection.md — which PDF library for which task, installation shortcuts.
• references/pdf-to-markdown.md — PDF → Markdown: decision tree (digital vs scanned), extraction recipe, pitfalls (multi-column, borderless tables, cross-page tables, headings), and why Markdown composition stays agent judgement.
• references/forms.md — AcroForm vs XFA, filling with pypdf, flattening, visual overlay fallback.
• references/weasyprint-setup.md — install platform notes, @page recipes, font embedding, page breaks.
• references/html-conversion.md — html2pdf.py deep dive: 10-step preprocessing pipeline, _NORMALIZE_CSS rules, reader-mode candidate list with body-ratio guard, render-time hardening (offline URL fetcher + SIGALRM watchdog), PDF outline (bookmarks) from headings, per-platform notes (Fern / Mintlify / GitBook / Confluence / Хабр / vc.ru), honest-scope limitations.
• scripts/md2pdf.py — Markdown → PDF via weasyprint + markdown2; mermaid blocks pre-rendered to PNG via mmdc.
• scripts/html2pdf.py — HTML → PDF via the same weasyprint pipeline; reuses md2pdf's default stylesheet (opt-out via --no-default-css).
• scripts/pdf_merge.py — bookmark-preserving merger via pypdf.
• scripts/pdf_split.py — range, per-page, or fixed-chunk splitter.
• scripts/pdf_watermark.py — text/image watermark overlay via reportlab + pypdf; per-mediabox overlay caching for heterogeneous decks; cross-7 same-path guard.
• scripts/pdf_fill_form.py — AcroForm inspect/extract/fill/flatten via pypdf; XFA forms detected and refused.
• scripts/preview.py — universal INPUT → PNG-grid renderer for .pdf (via Poppler) and .docx/.xlsx/.pptx (via LibreOffice + Poppler). Byte-identical across all four office skills.
• scripts/pdf_extract.py — dumps a PDF's per-page text + tables to structured JSON via pdfplumber, with scan detection (image-only document → exit 10) and opt-in artwork extraction (--extract-images DIR: rasters copied as stored + vector figures cropped via Poppler). A dump, not a Markdown converter.
• scripts/pdf_ocr.py — OCR a scanned PDF into a searchable PDF via ocrmypdf (default eng+rus); soft-optional engine (install.sh --with-ocr); imports _errors.py read-only (no cross-skill replication). S

…(truncated)