batch-fix-subtitles.md

Fix typos and re-sync subtitle timing in one pass

Auto-generated subtitles always need cleanup. Doing it in CapCut means clicking each line, fixing the text, then nudging the start point. With capcut-cli, one pipeline handles dozens of changes.

See what you've got

capcut texts ./project | jq -r '.[] | "\(.id[0:8])  \(.start_us / 1000000 | floor)s  \(.text)"'

a1b2c3d4  0s   Welcom to the chanel
e5f6g7h8  3s   today were lookng at
i9j0k1l2  6s   the new framwork release

Batch the fixes

capcut batch reads JSONL from stdin. One IO, one backup, one write.

capcut batch ./project <<'EOF'
{"cmd":"set-text","id":"a1b2c3d4","text":"Welcome to the channel"}
{"cmd":"set-text","id":"e5f6g7h8","text":"Today we're looking at"}
{"cmd":"set-text","id":"i9j0k1l2","text":"the new framework release"}
{"cmd":"shift-all","offset":"+0.3s","track":"text"}
EOF

shift-all with --track text nudges every subtitle by 300ms — useful when the speaker leads the captions.

Drive it from a spreadsheet

Export your fixes to a CSV (id,text), then convert to JSONL:

awk -F',' 'NR>1 {printf "{\"cmd\":\"set-text\",\"id\":\"%s\",\"text\":\"%s\"}\n", $1, $2}' fixes.csv \
  | capcut batch ./project

Drive it from an LLM

# extract → send to your LLM of choice → pipe back
capcut texts ./project \
  | your-llm-cli "Fix typos. Output JSONL of {cmd:'set-text',id,text}." \
  | capcut batch ./project

That's the whole loop. The CLI's JSON-in/JSON-out shape exists for exactly this.