MonkeeTools' hand-written PDF engine assembles files in about 130 lines, with WinAnsiEncoding as its honest tradeoff

The cross-reference table is where this design gets expensive. Each object is serialized with its byte offset recorded, and the trailer is written out of those recorded numbers ^[7][10]. Edit a string in the content stream and every offset after it moves. The object count is one line, `const N = 14 + (logoImg ? 1 : 0) + (payURL ? 1 : 0)`, because the logo image and the Pay now link annotation are optional, and the post says getting that number wrong corrupts the file, so it is derived and never hardcoded ^[9].

Work the loop out and the base document is thirteen numbered objects. N is written into the `xref 0 N` header and into `/Size`, and the fill loop runs from 1 to N minus 1, so 14 covers the free object 0 plus objects 1 through 13; with both optional objects present, N is 16 and there are 15 ^[10][11]. The post enumerates them: catalog, page tree, page, content stream, two fonts, font descriptors, font files, annotations ^[12].

Drawing is two closures. `txt()` pushes a begin-text, font select, size and color, move, show-string, end-text sequence; `line()` strokes a rule; the template renderers push hundreds of those, and the whole array becomes one content-stream object ^[8].

"The first time a real PDF reader opened the output without complaining, I celebrated more than I should admit," the author wrote ^[18].

Two subsetted Inter files, regular and bold, are embedded as FontFile2 streams with a hand-written font descriptor each, so the brand font does not fall back to Helvetica on systems that lack it ^[13]. The font dictionary declares `/FirstChar 32 /LastChar 255` and `/Encoding /WinAnsiEncoding` ^[14]. That line sets the product's language coverage. The post calls WinAnsiEncoding "the honest tradeoff at the heart of this engine": Latin-1 plus a slice of typographic characters, enough for every Western European language, and no emoji, CJK, Arabic or Cyrillic ^[15].

The escaper has four branches. Printable ASCII passes through, with backslash and parentheses escaped for PDF string syntax; anything under code point 256 becomes a three-digit octal escape; a lookup table maps the euro sign, ellipsis, curly quotes, en and em dashes and the trademark sign to their WinAnsi byte values; anything else is normalized NFKD with non-ASCII stripped, and if nothing survives it emits a question mark ^[16]. The code comment reads "Last resort: strip diacritics, else '?'. Never emit a broken byte." ^[17]

For the 300KB comparison to transfer to your app, your documents have to look like these: one page, text and stroked rules, an optional raster logo, one link annotation, Latin script ^[6][9][15]. The post does not state the size of its own generator, and it does not test what jsPDF or pdf-lib would do with a Cyrillic customer name ^[20]. Given that population, I would take the 130 lines. The day a name arrives in Japanese, the invoice prints a question mark ^[15][16].