DEV Community

dylan yu
dylan yu

Posted on

Three file jobs that keep breaking: image provenance, chat exports, and subtitle scripts

A lot of “AI tools” content is really about files. You generate or download something, then you need it in a format a teammate, a CMS, or a lawyer can actually use. Three jobs show up constantly in that gap:

  1. What metadata is still inside an image after you export it?
  2. How do you get a ChatGPT thread into Word without wrecking lists and code?
  3. How do you turn an SRT file into a document someone can comment on?

They look unrelated until you put them on the same desk. Below is a practical workflow I use, plus the three small browser tools I keep for the last mile.

1. Images: EXIF is not the whole story

Stripping “metadata” usually means EXIF: camera model, GPS, dates. That is not enough anymore.

  • IPTC still carries captions, credits, and sometimes keywords that survive a “Save for web” pass.
  • C2PA (Content Credentials) is a signed provenance box. Some generators attach it; some editors leave it; some CDNs strip it. You cannot guess from the pixels.

If you publish product shots, screenshots, or user uploads, inspect the output file, not the editor UI. A pipeline can remove EXIF and still ship a C2PA manifest, or the reverse.

I keep a short inspector/remover walkthrough here: Remove AI Meta. Use it as a hygiene check. It does not make a file anonymous, and it does not prove a picture was or was not generated. If the file is going into news, legal, or a client archive, run a second inspector on the bytes you actually attach.

Checklist I use:

  • Re-download the file you will send (not the preview in the tab).
  • Confirm the format (JPEG vs PNG vs WebP); manifests do not travel the same way.
  • If you must keep provenance (newsroom, brand kit), do not strip C2PA just because a blog said “remove AI meta.”

2. ChatGPT → Word: paste is not an export

A long ChatGPT thread is a structured conversation: turns, lists, fenced code, tables. Word expects paragraphs, styles, and headings. Clipboard paste maps none of that reliably. You get:

  • collapsed lists
  • code that wraps into prose
  • lost speaker turns
  • smart quotes that break snippets

For notes, a client recap, or an internal archive, you want a .docx you can search and edit. That is the job of ChatGPT to Word: browser-side export with readable blocks instead of a blob of text.

Operational rules, not marketing:

  • Redact first. API keys, customer names, and internal URLs should not hit Word’s autosave or email. The exporter is not an access-control layer.
  • Skim headings after export. If the thread mixed several topics, split the document before you share it.
  • Do not treat the .docx as a source of truth for code; copy snippets back from the original fence if you will compile them.

3. SRT → Word: players are not editors

SRT is a player format: index, timestamp, one or two lines of text. Translators, teachers, and producers need comments, tracked changes, and a printable script. Dumping SRT into a text editor leaves timestamps glued to dialogue; dumping it into Word by hand is slow and error-prone.

SRT to Word turns cues into an editable document and keeps timing blocks readable where the converter supports it.

Things that actually break batches:

  • Encoding (UTF-8 vs UTF-16 vs a Windows code page)
  • Cues that wrap mid-sentence
  • Overlapping timestamps from auto-caption tools

After conversion, I always read the first ten and last ten cues before sending the file. If those look right, the middle is usually fine; if they do not, stop and fix encoding instead of “cleaning in Word.”

How these three fit together

A typical week for me looks like: generate or receive assets → inspect what the file still contains → put the language into Word so other people can work.

Job Pain File you want
Image / screenshot Hidden provenance or leftover IPTC A clean (or intentionally signed) image
Chat session Unreadable paste .docx with structure
Subtitles SRT that nobody can comment on .docx script

None of these replace a DAM, a subtitle suite, or a records policy. They are the 10-minute last mile so you are not fighting the wrong format.

If you only bookmark one starting point for “browser tools grouped by job,” I keep a broader index at Web Tools List—but the three links above are the ones I actually open for this file workflow.

Top comments (0)