SubLime Captions

Interview videos by my friend Frank Yang, a YouTuber, took longer to caption than to edit. SubLime Captions is the tool I built for that job: upload a subtitle file, tell the AI what the video is about, and review what it proofread or translated. I did the research, interaction design, visual design and branding, and I built and shipped the product myself with AI coding tools. On a 15 to 20-minute interview, the work went from 8 to 12 hours to about 1 hour.

Context

In 2024 and 2025 I helped that friend shoot interview documentaries. Each video ran 15 to 20 minutes and produced 800 to 1,000+ subtitle lines from automatic speech recognition (ASR).

The ASR output was never clean. It got names and places wrong, misheard words, scattered punctuation, and kept every “um” and false start. The errors followed no pattern, so find and replace couldn’t fix them.

General-purpose AI didn’t solve it either. ChatGPT and Gemini couldn’t take a whole subtitle file in one go and return it with the timestamps intact. So the work stayed manual: read the file line by line and fix errors by hand, then, for translation, copy each line into a translator and paste the result back. One video took about a full working day.

AI was already good enough to fix the text. What was missing was a workflow that could feed a long file to AI in pieces, keep the context, and write the results back safely, with a human still in charge.

Case 1 · I replaced a four-field form with one prompt box

My first prototype, generated in Google AI Studio, asked for context in four fields: Speaker Names, Topic, Keywords and Additional Context. On my friend’s real file the corrections were good enough to prove the idea. But filling it in felt like paperwork. Before typing anything, you had to decide which box each fact belonged in.

One box, and you can leave it empty

Now a creator can write “This interview was filmed in Paris. The speaker is a designer. ‘Apple’ means the company, not the fruit,” and the AI works out the rest. A hint menu next to the box shows first-time users what kind of context helps.

The instruction is also optional. Most AI tools wait for a prompt, but here you can press send right away and the AI proofreads from the subtitles alone. The colour of the send button’s icon shows whether custom instructions are included. Settings people rarely change (filtering filler words, stutters and profanity, and punctuation handling) went into a + menu, where ChatGPT and Gemini put theirs.

Pause

Case 2 · Every AI edit stays visible and can be undone

Compare view shows the original and the AI version side by side. Preview view shows only the final text, for a read-through. Status filters (All, Modified, Original) take you straight to the 200 lines the AI changed and skip the other 800. In translation mode the labels switch to “Translated”, so they always match the task.

Pause

Status you can see while scrolling

The line number doubles as status. It turns green when a line is processed and red when it fails, so problem lines stand out as you scroll past them.

Undo for one line or a whole paragraph

Revert and restore work on one line or on a selection, so a creator can reject a whole paragraph of translation and bring it back later. Manual edit covers the cases where neither version is right. Delete sits in the opposite corner from Save and Cancel, so a destructive action is never one slip away.

Pause

Case 3 · Selecting 200 lines takes one drag

Creators often want AI on one part of the video, like a single interview segment. In a 1,000-line list, that needs faster ways to select and to check what’s selected.

Drag, invert, and jump

Press on any checkbox and drag across the others to select or deselect a run of lines. Drag past the edge of the editor and the list scrolls, so the selection keeps going. Select all and invert handle “everything except this part” in two clicks.

Pause

The selection pill sits above the prompt box and shows what’s selected, for example “Lines 20–50”. Selections in a long file are often scattered, so clicking the pill jumps through each selected line in turn. Creators can check what the AI is about to touch without scrolling. Selecting a line also brings back the prompt box if it was hidden, because selecting usually means you’re about to act.

Pause

Search the way people remember

People remember a problem line as a phrase, as “line 120”, or as “around 3:20”. Search matches subtitle text, line numbers and timestamps, so all three work.

Case 4 · Proofreading and translation run in either order

When you translate after proofreading, the AI works from the corrected text and leaves the raw ASR behind. Each switch also saves the previous stage, and the download menu offers both the current and the previous version. A creator can export corrected English subtitles and the French translation from the same session.

Pause

Three controls that appear only for translation

Translation mode adds a source language (auto-detect by default, because interviews often mix languages), a target language and a style slider. Faithful stays close to the original wording. Natural reads smoothly and keeps the meaning. Localized uses the idioms and phrasing a native viewer would expect, which suits social video.

Pause

Principles

Three principles shaped the decisions above.

1. [01]

   Say it, don't fill it in

   Context goes in as plain language, not as form fields (Case 1).

2. [02]

   AI proposes, the creator decides

   Every AI change is visible, reversible and editable (Case 2).

3. [03]

   Built for 1,000 lines, not 10

   Every interaction has to hold up on a long file (Cases 3 and 4).

A lime-green identity that stays out of the way

“SubLime” combines subtitle and lime, and sublime also means refined, which is what the tool does to captions. The logo is a lime slice with a play button in it. I chose lime green (#32CD32) and a few neighbouring greens because they feel fresh and energetic without looking cold.

The colour only appears where something needs attention: selected checkboxes, hover states, the soft glow around the editor while AI is working, and the progress bar. Subtitle text stays neutral, because reading it is the job. As the product grew, I turned these choices into a token-based design system with semantic colours for light and dark themes.

How I built it

The AI Studio prototype took one prompt and a few minutes to generate. Turning it into a complete product took about six months, from December 2025 to May 2026, working in Antigravity and Codex.

Most of the interaction design happened in the running product. I described a behaviour in plain language, let the AI implement it, used it on a real subtitle file, and refined it. Some details only show their problems once you can use them. Should the prompt box hide while you read? How fast should the selection pill jump? I couldn’t have answered either from a static mockup. For layouts that were hard to describe, such as where a menu should open, I sketched quick wireframes in Figma and handed them to the AI.

That speed only helps if the instructions are precise. “Make this button look better” got me nothing useful. What worked was a description of every state:

Light grey background, larger corner radius. On hover, fill with the lime theme colour. When the prompt box has text, turn the icon lime and add a soft shadow so it lifts slightly. Keep the transition subtle, around 0.2 seconds.

On a team I’d still use Figma to explore options and align with engineers. As a solo builder, the fastest way to judge an interaction was to use it.

The hardest part was one creators never see: splitting a long file into batches. Batch size adapts to line length, several batches run in parallel, and each batch carries the lines before it as read-only context, so the AI doesn’t lose track of the conversation.

Outcome

SubLime Captions is live at sublimecaptions.com and free to use. Core development wrapped up in May 2026. Every request calls a paid AI model, and I pay for the API myself, so I haven’t promoted it. I use it as my own tool and as practice in product design. Even so, about 180 people have found the site on their own and visited it about 250 times (as of October 2026). I don’t have data yet on how well it works for anyone other than the creator it was built for.

For that creator, proofreading and translating a 15 to 20-minute interview of 800 to 1,000 lines used to take 8 to 12 hours. With SubLime Captions it takes about 1 hour: the AI pass takes 2 to 5 minutes, and the rest is the creator’s review.

What changed after launch

A senior product manager reviewed the product and pointed out a risk: when a creator lists a name or term, the AI may treat it as something to insert, when it is only a spelling reference. In testing, the model sometimes did exactly that. I rewrote the prompt so user notes apply only where relevant, treated bare terms as spelling references, and added a regression test.

Because I pay for every request, I set a daily AI allowance per IP address. The FAQ used to say only “free during beta”. It now explains the limit in units creators understand: roughly 15,000 to 80,000 subtitle characters a day, resetting at midnight UTC.

I’ve also moved through several Gemini Flash versions and tuned the prompts along the way. The interface didn’t need to change, because review and control never depended on the model being perfect.

Reflection

In an AI product, much of the experience lives in review and control. The AI pass on a 1,000-line file takes minutes. Most of that hour, and most of my design work, went into how a creator sees what changed and decides what to keep.

Building with real data from day one changed how I validate design. I found problems by using the product on real files, which a mockup could not show me. Outside review still matters when you build alone: one conversation with a product manager surfaced a failure mode I hadn’t tested for.

What’s next

My research came from one creator’s workflow. Next I want to test with a wider group of creators, especially non-English and multilingual channels. I’d also add privacy-respecting analytics, such as how often AI edits are kept or reverted. That rate would be the clearest measure of trust.