SUBTXT 2

THE IDEADEVELOPMENT

A system that captions video automatically

SUBTXT began at a 24-hour hackathon where we were asked to pair a generative video API with a language model. We built a system that automatically captions unexplained references and concepts, so casual viewers can better follow expert conversations.

3 research artifacts available on desktop

Focus Mode's design suggested a narrow framing of the problem - treating distraciton as something external, to be managed by controlling the environment.

To better understand focus and distraction, I turned to behavioural frameworks. In Indistractable, Nir Eyal writes:

While we love to blame external triggers...most of our distractions begin from within

This reinforced my suspicion that we were conceiving the problem too narrowly and provided inspiration for concept testing.

This misalignment wasn't just theoretical, it was reflected in the tools available.

Expert conversations often contain references casual viewers do not understand, which go unexplained by the hosts.

A product that automatically detects these moments and generates short, contextual captions to explain them.

What was said

What SUBTXT captions

HUMAN–AI INTERACTIONDEVELOPMENT

Designing the creator experience

LLMs and video-generation tools made finding references and writing captions relatively straightforward. The harder challenge was the creator experience around that automation. I defined the division of labour, what SUBTXT proposes and what the creator decides, then turned it into a step-by-step workflow.

🤝

AI leads, creators decide

🪄

Make it easy for non-editors

💎

Make value visible at every step

The Workflow

Instead of exposing every decision in a Premiere Pro-style canvas, I separated them into focused steps that non-editors could work through progressively.

Division of labour

Rather than a one-shot handoff, I introduced review points into the automated process: AI recommends and generates; creators approve, adjust or override.

The workflow gives creators clear points to review and shape the AI’s work, without the complexity of a traditional editing canvas.

PROTOTYING IN CODEDEVELOPMENT

Designing the "Curate" step

The Curate step looked straightforward on paper, but in practice AI recommendations, creator decisions, and caption timing affected one another in unexpected ways. Rapid prototypes with Claude made these interactions visible early.

See how the curate step evolved →

The demo landing page communicated the concept through a working example.

Caption settings gave creators control over the AI's output.

Users can browse and interrogate moments

System recommends captions to includeents

Moments that clash are flagged

Requirements brief for Claude

Early prototype

Prototype snapshots

See how the Curate step evolved · 3 min →

Key Design Decisions

‍

This walkthrough above focuses on what users experience. The notes below explain the reasoning behind the most important design decisions. (Coming soon!)

Users had already expressed interest in accountability-based support
It addressed a clear gap in existing focus tools
It aligned with behavioural research on how focus actually breaks down
Early thinking suggested the technical implementation was feasible

The Redesigned Experience

The prototype proved the concept, but it still reflected a demo rather than a product. I continued developing the concept independently, prioritising user control and clarity.

Users had already expressed interest in accountability-based support
It addressed a clear gap in existing focus tools
It aligned with behavioural research on how focus actually breaks down
Early thinking suggested the technical implementation was feasible

LANDING PAGE

Designing the AI's behaviour

The product relied on a multi-step pipeline: Gemini identified unfamiliar references before generating contextual captions, which Hera rendered directly into the video. I designed the editorial, writing and presentation rules that guided each stage of that process.

Video in

Flag moments worth explaining

Editorial criteria ›

Generate multiple caption briefs

Caption brief ›

Render captions, consistently

Presentation rules ›

Video out

What we shipped in 24 hours

With limited time to design the interface, I focused on the two capabilities that best demonstrated the product's value: explaining unfamiliar concepts and giving creators control over the AI's output.

The demo landing page communicated the concept through a working example.

Caption settings gave creators control over the AI's output.

The demo landing page communicated the concept through a working example.

Caption settings gave creators control over the AI's output.

Evolving the interaction model

After the event, I continued developing the concept independently. While the prototype demonstrated the technology, it lacked the level of control needed for a professional creative workflow. Rather than asking creators to trust a one-shot AI output, I redesigned the experience around a series of editorial decisions.

SUBTXT — Four steps

STEP 01

⬆️ Upload

Get the video in, and confirm it's the right one.

Provide the video
Confirm it's the right one
System reads market, language, lengthAI

STEP 02

✍️ Curate

Decide which moments make the cut, at a sane volume and rhythm.

See everything found
Understand each reference
Understand why it was flagged
Include or exclude
Judge total volume
Judge cadence
Preview in context

STEP 03

🎬 Edit

Refine each caption — copy, imagery, timing and position.

Edit title and explanation
Change icon
Swap or remove image
Adjust timing and duration
Set on-screen position
Space out flagged clusters

STEP 04

⬇️ Export

Apply brand, watch it through, deliver the finished video.

Apply brand font and colours
Final watch-through
Export and download

A collaborative interaction model replaces the one-shot approach from the demo, offering a better experinece with more confidence and the ability to control alll the key asepcts of the product's output.

The recommendation logic

What the prototype revealed

Weighing two fixes

Testing the second fix

The Solution