Claude Opus 4.5 in Action: One Skill That Turns PDFs into Interactive Animated E-Books

From One-Shot Video Generation to Interactive E-Books

The multimodal generation capabilities of large language models are evolving at a visible pace. Claude Opus 4.5's ability to "one-shot a video" has already sparked discussion in the developer community — but one developer decided to push that capability further, applying it to a more complex use case: automatically transforming a static PDF into an animated, narrated, interactive web-based e-book complete with quizzes.

The project is called Papermorph. It started as a Skill (a Claude capability extension) with the goal of converting a book into a series of educational videos. As the author continued refining it, the project evolved into a tool that generates complete "Web Books" — not linear videos, but freely explorable animated lessons with narration and interactive quizzes.

In terms of product design, this represents a notable trend in AI content generation: moving from "generate a piece of media" toward "generate a fully interactive experience."

Breaking Down Papermorph's Workflow

The author laid out a clear processing pipeline, which is itself worth studying:

PDF → Book Plan → Storyboards → Narration → Animation & Quizzes → Web Book

Why This Pipeline Matters

Each step maps to a real task in content creation. Traditionally, adapting a book into an interactive course requires instructional designers, scriptwriters, voice actors, animators, and front-end engineers working in parallel. Papermorph compresses all of these stages into a single automated pipeline, driven by nothing more than Opus 4.5 and this one Skill.

One detail worth highlighting: as the author notes, the entire system currently does not use any image generation models. The language model alone parses the PDF and organizes the content — and already produces results that are, in the author's words, "surprisingly good." This suggests that the current generation of large models has matured considerably in structured understanding of long documents and automated narrative planning.

---

What is a Skill? A Skill is a capability encapsulation mechanism in the Claude ecosystem that allows developers to package a multi-step AI workflow into a reusable module. Compared to raw API calls, a Skill is more like a "preconfigured expert role" — it carries its own system prompts, tool-calling logic, and context management strategy. In Papermorph's case, the Skill coordinates every step from PDF parsing to final web page generation, so developers don't need to manually design prompt chains on each call. This encapsulation approach is analogous to function abstraction in software engineering — hiding complex processes behind a unified interface to lower the barrier to reuse.

---

How Far Can a Pure Language Model Go Without Image Models?

This is the most technically interesting aspect of the project. At a time when many assume "visual content requires diffusion models," Papermorph demonstrates that the planning and organizational capabilities of a language model alone can drive the generation logic for animation and interactivity.

The content of animations and storyboards is determined by the language model — it breaks down concepts, designs the pacing of explanations, and structures quizzes — while the presentation layer is implemented through code (web animations) rather than rendered pixel images. This approach of "language model as director, code as renderer" is more controllable than calling an image model directly, costs less, and is better suited to the accuracy requirements of educational content.

That said, the author acknowledges there are limits to pure text-driven generation. The next planned step is precisely to introduce image models for storyboarding, enabling the tool to handle content types that rely heavily on visual expression — such as picture books and humanities documentaries.

---

A note on diffusion models: The "diffusion model" mentioned here refers to the underlying architecture behind mainstream image generation technologies today, including Stable Diffusion, DALL·E, and Midjourney. The core idea is to iteratively remove noise to restore or create pixel images. Papermorph sidesteps this entirely by having the language model output structured HTML/CSS/JavaScript animation code — effectively delegating "visual presentation" to the browser's rendering engine rather than a pixel-generation model. This trade-off has clear advantages: the animation content is fully editable, semantically consistent with the text, and free from the visual hallucinations common in diffusion models (such as abnormal numbers of fingers on hands). The limitation is that pure code-based animations struggle to render photorealistic imagery, which constrains expressive power for content types that require real visual assets, such as historical documentaries.

---

Open Source Under MIT License

Another strong point of the project is its openness. The author provides two entry points:

• Online bookshelf (continuously updated): https://papermorph.diamonddoge.org/ — browse completed Web Book outputs directly.
• GitHub repository: https://github.com/DozenTwelve/Papermorph — released under the MIT License, meaning developers are free to use, modify, and even commercialize it.

The permissive MIT license is very community-friendly and lowers the barrier for others to integrate image models or swap in different large models on top of this foundation.

Observations and Takeaways

First, the value of Skills as capability encapsulation is becoming clear. Packaging a complex workflow into a reusable Skill allows a single developer to independently accomplish what once required an entire team. This is an important direction for the engineering of generative AI.

Second, educational content may be one of the best use cases for multimodal generation. Transforming dry PDF textbooks into interactive courses with quizzes addresses a real pain point. Educational content also demands less "wow factor" than entertainment and places more emphasis on structural clarity and factual accuracy — a better fit for current model strengths.

Third, projects like this are still in early stages. The current results come primarily from a single developer working with a single model. Large-scale validation is lacking. The stability of generation quality across different PDF types, the accuracy of long-document processing, and the pedagogical effectiveness of the auto-generated quizzes all require more real-world testing.

For developers and content creators following AI application development, Papermorph offers a useful case study: it demonstrates a complete approach to decomposing frontier model capabilities into a clear pipeline and packaging them into a reusable tool.