Claude Code: create Vietnamese captions and check errors

How can Claude Code help create Vietnamese captions?

Claude Code can coordinate transcription, timing data and caption rendering in a suitable environment. You still need readable audio, speech recognition or an approved transcript, and rendering tools. A prompt alone does not guarantee accurate word-level captions for every video.

Remotion documents SRT/VTT import, on-video caption display and separate SRT export. These workflow capabilities do not prove a particular transcript is correct. BHKD authored this step-by-step guide for beginners; it is not a report of an executed video test.

1. Choose the output: burned-in captions or SRT?

Burned-in captions are rendered into the image and cannot be switched off separately. SRT stores subtitle text and timing for tools or platforms that support it. SRT does not inherently preserve the complete animated appearance of a rendered video. For reuse, retain a clean video, a captioned export and the approved transcript.

Output prompt: Inspect [video path] and available tools. I need an MP4 with Vietnamese captions plus a separate SRT; preserve the source. Explain the steps, transcription tool, timestamp support and any costs. Use existing data only; do not upload to external services or use paid APIs without permission. Disclose unreadable audio before proceeding.

2. Transcribe with a protected terminology list

Prepare confirmed names, brands, specialist terms and spellings. This list helps detect errors but does not justify replacing every similar-sounding word. Recheck amounts, percentages, dates and negations; a missing no can reverse meaning.

Transcription prompt: Create a timestamped transcript from [video] using the agreed tool. Use [terminology file] to check spelling without inventing unclear speech. Mark uncertain passages [unclear] with timecodes for review. Preserve negations and qualifications. Provide text for approval before effects or full rendering.

3. Correct text and retain a change log

Keep machine and approved transcripts separate. For important edits, record timecode, before/after text and reason. Correcting spelling differs from rewriting a statement. Any readability shortening must preserve meaning and must not misrepresent new wording as a quotation. Assign a reviewer for sensitive material.

Text-review prompt: Compare the transcript with confirmed names and my corrections. Propose changes in a table rather than bulk-editing uncertain words. Flag figures and negations for replay. Create an approved version only after confirmation; do not label unreviewed text approved.

4. Synchronize captions after the edit is locked

Removing footage or changing speed after caption generation can invalidate timing. Lock the edit used for timestamps or update timing against the final timeline. Distinguish a constant offset from increasing drift; one fixed correction is not a universal solution.

Synchronization prompt: Check the beginning, middle and end of final [video] against the approved transcript. Report constant offset versus progressive drift and identify passages to fix. Use word-level animation only with reliable timestamps; otherwise propose phrase captions. Do not fabricate timings for visual completeness.

5. Design readable text without constant animation

For vertical video, begin with at most two lines, meaningful line breaks, Vietnamese-capable type and clear contrast. This is an exercise starting point, not a universal platform rule. Check placement in the actual publishing interface to avoid controls; a still preview alone cannot guarantee safe areas.

Design prompt: Create a ten-second draft with [licensed font], white type, restrained blue highlights and a navy background only where contrast requires it. Avoid icons and repetitive effects that add no meaning. Preserve Vietnamese marks, avoid faces and improve line breaks to prevent stranded final words where possible. Export three frames and a draft for phone review.

6. Export files and inspect the final version

Deliver the source, captioned MP4, SRT, approved transcript and unresolved-issue notes. Match the SRT to the correct edit, not an earlier cut. If a platform adds automatic captions, check for double captions over an already burned-in version.

Acceptance prompt: Inspect exported MP4 and SRT for existence, matching duration/version and unintended timing overlaps. State the tools used and where manual replay is required. Do not claim all words are accurate merely because rendering succeeded. Provide a delivery table with approved or pending status.

A 60-second exercise: prioritize accuracy

Choose clear speech with limited background music and a brand name or figure. Create phrase captions and, only if timing allows, word-level captions. Compare readability, correction count, processing time and distraction. More animation need not be the better option.

Business lesson: captions are part of brand communication and content accessibility, not just decoration. Approved text, reliable timing and final review are often more valuable than applying new effects to every video.

Suggested actions