How to Edit Auto-Generated Captions Before You Post

X has rebuilt its video composer around one change: captions you can now correct before you post. If you have ever wondered how to edit auto generated captions without fighting each app separately, this is a good moment to settle on a habit. Here is what X shipped, why automatic captions get the same kinds of words wrong everywhere, and a check-and-fix pass that works before any video goes up.

What X shipped

X developer Allegra Jacchia announced an updated version of X's native video editor in a post on X, reported by Social Media Today on 4 October and Social Samosa on 5 October. The composer has a more focused layout, and its main change is captioning:

• Captions are editable in the composer. You can correct generated captions by hand before posting. Jacchia called it "one of our most requested features".
• Grok can correct them for you. You can ask X's Grok to fix the captions instead of retyping words yourself.
• Grok can translate them. The same tools can put captions into another language.

What the announcement does not say is just as useful. Neither report names the regions, account types or apps that get the new composer, or whether it is fully rolled out. It builds on a July update that added multi-language caption tools, and Jacchia said more improvements are coming. If you do not see editable captions in your composer yet, that is consistent with everything published so far.

The bigger signal is the direction. Instagram's new Reels guide tells creators to review its automatic captions for accuracy, which we covered in how to add captions to Instagram Reels. Now X has made the same correction a step in the posting flow. Platforms have stopped treating automatic captions as finished.

Why auto captions get the same words wrong

Speech-to-text is very good at ordinary sentences and predictably weak in the same few places. Knowing where to look turns a full proofread into a two-minute scan.

• Names. People, places, products and channels. A speech model has to guess at a word it has rarely heard, and it guesses something common that sounds similar.
• Jargon and brand terms. The more specialist your video, the more of its key words sit outside everyday speech. This is exactly where your video is most specific and the model is least sure.
• Numbers. Prices, dates, model numbers and percentages. "Fifteen" and "fifty" are close, and a wrong number is worse than a missing one.
• Words that sound alike. "Their" and "there", "affect" and "effect". The audio is identical, so only the sentence tells them apart.
• Overlapping voices and music. When two people talk at once, or a track sits under the voice, words get dropped or merged.
• Short clips in a language the model has to guess. With only a few seconds of audio, automatic language detection has little to go on.

None of this is a fault of any one platform. It is how these models behave, which is why the fix is a habit, not a feature.

A check-and-fix pass for any platform

Run this once, on the final cut, whatever tool made the captions.

1. Read the transcript with the sound on. Scan the text, not the video. Your eye catches a wrong word in a list far faster than in a moving caption.
2. Search for every name and number. Check each one against what was actually said. These are the errors viewers notice and screenshot.
3. Fix the first caption first. It is on screen when the viewer decides whether to stay.
4. Check the timing at the cuts. After a trim, make sure no caption runs past the end of its sentence or appears before the speaker starts.
5. Watch it once on mute. Captions exist for the viewer with the sound off. If the video makes sense silent, you are done.

Fixing captions in flshy before you upload

If you caption in your editor rather than in each app, you make every correction once and the fixed captions go wherever the file goes. In flshy the whole pass runs free in the browser, and the transcription happens on your own machine. The full walkthrough is in add captions to a video; the correction-specific parts are these.

1. Cut the clip to length first. Caption cues sit on their own track and do not move with a silence removal pass, so captioning after the cutting is what saves transcribing twice.
2. Name the language if the clip is short. In the Captions panel, the language picker offers auto detect and 14 named languages. Naming it yourself helps most on short clips, where there is not much audio to judge from.
3. Generate, then put the cues on the timeline. Press Generate captions, then press the button under the transcript that reads Add 24 as Editable Text, or whatever your count is. Until you press it, there is nothing on the timeline to correct.
4. Read the Transcript tab. This is step one of the pass above: the words as a list. Check names and jargon here first.
5. Retype, restyle or drag. Every cue is an ordinary editable text clip on a Captions track. Retype a misheard name, restyle a single line, or drag a cue that lands early.
6. Let trims tidy up after you. Cues stay attached to the clip they came from. Trim the clip and its cues are shortened to match, and any cue left entirely outside the new edge is removed.
7. Export with the captions burned in. H.264 MP4, up to 4K, no watermark.

Platform captions or captions in the file

Both are reasonable, and the right answer depends on where the video lives.

• Posting to one platform only? Its own caption tools, now including X's editable captions, keep everything in one place.
• Posting the same cut to several? Captions burned into the file mean one correction pass covers X, Reels, Shorts and TikTok. If you burn them in, leave the platform's automatic captions off for that post, or viewers get the same words twice.
• Need the viewer to be able to switch them off? Burned-in captions cannot be turned off. That is the one real trade.

What to watch next

The open questions are where X's new composer is available, which languages its translation covers, and whether other platforms follow with in-composer correction of their own. We will cover them here with the sources linked. Until then, the captions guide behind this post, and the rest of the guides, are where the detail lives.