pbpaste-md — paste clipboard as clean Markdown on macOS. Converts RTF/HTML (Word, Pages, Confluence, Teams, browsers) via pandoc, stripping non-semantic spans, div wrappers, and Confluence prosemirror artifacts. Falls back to plain text.

pbpaste-md

#!/usr/bin/env bash
# pbpaste-md — paste the clipboard as Markdown, preserving formatting.
#
# Strategy:
#   1. If the clipboard has RTF (Word, Pages, TextEdit, Apple Mail, …) and
#      `textutil` is available, convert RTF → HTML via textutil and then to
#      Markdown via pandoc. RTF preserves list semantics that Word's HTML
#      flavor mangles (it fakes bullets with <span style="font-family:Symbol">).
#   2. Else if the clipboard has HTML (Teams, Confluence, browsers, …),
#      convert directly with pandoc.
#   3. Else, fall back to plain-text pbpaste.
#
# Both HTML inputs go through clean_html() before pandoc to remove non-
# semantic wrappers that GFM would otherwise pass through as raw HTML.
#
# Usage:
#   pbpaste-md              # markdown (or plain text fallback)
#   pbpaste-md -r           # raw HTML (RTF is converted to HTML if needed)
#   pbpaste-md -t           # force plain text (same as pbpaste)
#   pbpaste-md -h           # help

set -euo pipefail

usage() {
  sed -n '2,21p' "$0" | sed 's/^# \{0,1\}//'
}

mode=md
while [ $# -gt 0 ]; do
  case "$1" in
    -r|--raw|--html) mode=html ;;
    -t|--text)       mode=text ;;
    -h|--help)       usage; exit 0 ;;
    *) echo "pbpaste-md: unknown option: $1" >&2; usage >&2; exit 2 ;;
  esac
  shift
done

clipboard_info=$(osascript -e 'clipboard info' 2>/dev/null || true)
has_html() { [[ "$clipboard_info" == *"class HTML"* ]]; }
has_rtf()  { [[ "$clipboard_info" == *"class RTF"*  ]]; }

dump() {
  # dump <four-char-code> <out-file>
  # Writes the requested clipboard flavor to $2 via AppleScript.
  osascript \
    -e "set f to (open for access (POSIX file \"$2\") with write permission)" \
    -e "write (the clipboard as «class $1») to f" \
    -e 'close access f' >/dev/null
}

# clean_html
# Reads HTML on stdin, writes cleaned HTML on stdout. Removes non-semantic
# wrappers that confuse pandoc's GFM output:
#   1. textutil heading paragraphs (large font, e.g. 16px) → <h2>.
#   2. Confluence/prosemirror inline code: <span class="code" …>X</span> →
#      <code>X</code>.
#   3. Confluence emoji placeholders: <span data-emoji-text="X" …></span> → X.
#   3b. Bold spans (style="font-weight:600|700|bold") → <strong>, so emphasis
#      survives the blanket span strip below.
#   4. Strip ALL remaining <span> open AND close tags independently (color
#      marks, mention pills, prosemirror marks, Apple-converted-space, sN, and
#      Microsoft Loop/Scriptor's nested attribution wrappers, etc.) — keeping
#      content. Stripping open/close separately (rather than as matched pairs)
#      is essential: Loop nests spans many levels deep, and a paired
#      non-greedy match leaves stray outer open tags behind.
#   5. Strip <div …> and </div> wrapper tags (preserving inner content) —
#      Confluence's fabric-editor layout divs and textutil's loose-text divs
#      both belong here. Pandoc otherwise passes them through verbatim.
#   6. Drop the "code-block" class on <pre>/<code> so pandoc doesn't emit it
#      as a code-fence info string.
#   7. Trim whitespace inside <p> after the rewrites.
#   8. Group runs of "<p>• text</p>" faux-bullet paragraphs into <ul>.
clean_html() {
  perl -CSD -0777 -pe '
    # Drop HTML comments (e.g. Scriptor'\''s <!--ScriptorStartFragment-->).
    s{<!--.*?-->}{}gs;

    my %heading;
    if (/<style[^>]*>(.*?)<\/style>/s) {
      my $css = $1;
      while ($css =~ /p\.([A-Za-z0-9_-]+)\s*\{[^}]*font:\s*([0-9.]+)px/g) {
        $heading{$1} = 1 if $2 >= 14;
      }
    }
    for my $cls (keys %heading) {
      s{<p class="\Q$cls\E">(.*?)</p>}{<h2>$1</h2>}gs;
    }

    # Inline code spans (Confluence prosemirror): match a <span> whose class
    # attribute is exactly "code" or starts with "code ".
    s{<span\b([^>]*\bclass="(?:code|code\s[^"]*)"[^>]*)>(.*?)</span>}{<code>$2</code>}gs;

    # Strip the "code-block" class on <pre>/<code> so it isn'\''t turned into
    # a code-fence language by pandoc.
    s{(<(?:pre|code)\b[^>]*\bclass=")([^"]*)(\b)code-block\b}{$1$2$3}gs;

    # Emoji placeholder spans (Confluence): extract the emoji character.
    # Must run before the general span strip so the text value is recovered.
    s{<span\b[^>]*\bdata-emoji-text="([^"]*)"[^>]*>.*?</span>}{$1}gs;

    # Bold spans → <strong> so emphasis survives the blanket strip below.
    # Matches font-weight: bold | 600 | 700 | 800 | 900.
    s{<span\b[^>]*\bstyle="[^"]*font-weight:\s*(?:bold|[6-9]00)[^"]*"[^>]*>(.*?)</span>}{<strong>$1</strong>}gs;

    # Strip ALL remaining <span> open AND close tags independently, keeping
    # inner content. Done separately (not as matched pairs) so deeply nested
    # wrappers — e.g. Microsoft Loop/Scriptor attribution spans — are removed
    # cleanly instead of leaving orphaned outer open tags behind.
    s{</?span\b[^>]*>}{}gs;

    # Reduce <a> tags to just their href. Pandoc passes anchors through as raw
    # HTML when they carry attributes it does not recognise (e.g. Microsoft
    # Loop'\''s unfurl / propertyattribution); with href alone it emits a clean
    # [text](url) link. Anchors with no href (bookmarks) lose their tag.
    s{<a\b[^>]*?\b(href="[^"]*")[^>]*>}{<a $1>}gs;
    s{<a\b(?![^>]*\bhref=)[^>]*>(.*?)</a>}{$1}gs;

    # Drop <div …> and </div> wrappers entirely; keep their content. Pandoc
    # otherwise re-emits them as raw HTML in GFM output.
    s{</?div\b[^>]*>}{}gs;

    # Trim whitespace that the rewrites left at the start/end of <p>.
    s{(<p[^>]*>)\s+}{$1}gs;
    s{\s+(</p>)}{$1}gs;

    # Group consecutive faux-bullet "<p>• text</p>" paragraphs into a <ul>.
    s{((?:<p[^>]*>\s*[\x{2022}\x{00B7}\x{25E6}\x{25AA}\x{25CF}]\s*.*?</p>\s*)+)}{
      my $block = $1;
      my @items;
      while ($block =~ m{<p[^>]*>\s*[\x{2022}\x{00B7}\x{25E6}\x{25AA}\x{25CF}]\s*(.*?)</p>}gs) {
        my $item = $1;
        $item =~ s/^\s+|\s+$//g;
        push @items, "  <li>$item</li>";
      }
      "<ul>\n" . join("\n", @items) . "\n</ul>\n"
    }gse;
  '
}

# html_to_md <html-path>
# clean_html → pandoc → minor whitespace cleanup.
html_to_md() {
  clean_html < "$1" \
    | pandoc -f html -t gfm --wrap=none \
    | awk 'BEGIN{blank=0} /^[[:space:]]*$/{blank++; if(blank==1) print ""; next} {blank=0; print}'
}

case "$mode" in
  text)
    exec pbpaste
    ;;

  html)
    tmp=$(mktemp -d -t pbpaste-md.XXXXXX)
    trap 'rm -rf "$tmp"' EXIT
    if has_html; then
      dump 'HTML' "$tmp/c.html"
      clean_html < "$tmp/c.html"
    elif has_rtf && command -v textutil >/dev/null 2>&1; then
      dump 'RTF ' "$tmp/c.rtf"
      textutil -convert html "$tmp/c.rtf" -stdout 2>/dev/null | clean_html
    else
      echo "pbpaste-md: no HTML or RTF on clipboard" >&2
      exit 1
    fi
    ;;

  md)
    if ! command -v pandoc >/dev/null 2>&1; then
      echo "pbpaste-md: pandoc not on PATH; install with 'brew install pandoc'. Falling back to plain text." >&2
      exec pbpaste
    fi

    tmp=$(mktemp -d -t pbpaste-md.XXXXXX)
    trap 'rm -rf "$tmp"' EXIT

    # Prefer RTF when present — Word's HTML mangles list structure into
    # styled spans, but its RTF carries real list metadata that textutil
    # turns into proper <ul><li>.
    if has_rtf && command -v textutil >/dev/null 2>&1; then
      dump 'RTF ' "$tmp/c.rtf"
      textutil -convert html "$tmp/c.rtf" -stdout 2>/dev/null > "$tmp/c.html"
      html_to_md "$tmp/c.html"
    elif has_html; then
      dump 'HTML' "$tmp/c.html"
      html_to_md "$tmp/c.html"
    else
      exec pbpaste
    fi
    ;;
esac