B[B]L[L]U[U]E[E]T[T]F[F]U[U]L[L]

The idea for Bluetful came from trying to make an AI short film. That process involved writing a script, planning shots, defining characters, generating scenes and images, and finally making video.

I used ChatGPT, Prompt Hub, Infinite Canvas, Jimeng, Lib TV, Xiaoyunque and other tools. Each handled one part of the work, but together they left the process fragmented. A script might live in a chat, character references in a local folder, and prompts on another platform.

Regenerating a shot, adjusting a character or reusing a prompt meant searching across several apps. I began asking whether the process could live in one place, where relationships between materials survived and finished work could remain useful. That question became Bluetful's starting point.

Canvas was the first core module. I built the basics: panning, marquee selection, multiselection, dragging nodes, connections and a toolbar. Then I began integrating AI generation into the node system.

A node can hold more than an image or text. It can also carry the prompt, model settings, parameters and context used to create a result. Different content types can be organized and connected into a larger workflow.

I later explored more specialized image features, including relighting, multi-angle generation, panoramas and a director's desk. The canvas grew from a content organizer toward a workspace for multimodal creation.

One canvas, more possibilities

I want the canvas to support text, images, video and audio, as well as documents, presentations and more complex interactive work in the future. Its value lies not only in infinite space, but in the ability to expand to new kinds of creation.

As the canvas, asset library and prompt modules took shape, I wondered whether people should still have to open every module and find every command themselves. Bluetful therefore began adding an AI Agent.

It started with conversation and tool calls, then expanded toward Skills, model routing, session management and operations on some workspace resources. The goal is for the Agent to understand a task and, with appropriate authorization, use the workspace's existing capabilities. Someone might ask it to find an asset, organize content or run a Skill.

Control matters

Generating content and performing actions carry different risks. Reading material is different from deleting it, and irreversible operations need a different kind of confirmation. I introduced tiered confirmation for tool execution and continue exploring the balance between efficiency and user control. The Agent should reduce repetitive work while keeping important actions understandable and controllable.

A refined prompt or a repeatable procedure can be more valuable than a single generated image or document. Bluetful therefore developed separate management for Prompts and Skills.

Prompts store reusable instructions. Skills carry more structured capabilities that an Agent can use for particular methods and tools. Both can be managed, searched, reused and brought into Agent workflows. I want to preserve creative knowledge alongside the output.

Bluetful began as a personal tool. As its Agent, canvas generation and asset management evolved, I began turning it into a SaaS product with accounts, cloud storage and commercial capabilities.

Different models have different capabilities, parameters and costs. How should a person choose one? How should a task's cost be estimated in advance? What happens to reserved credits if it fails? I introduced a model registry so canvas and Agent features can share one description of model capabilities.

I also designed a Credits system for model usage, subscription benefits, balances, task budgets and consumption records. It has been rebuilt several times to handle reservations, final settlement, refunds and credits with different expiration dates. This work pushed me to think about how the whole AI service runs, beyond the generation interface.

July 2026 — Start with the canvas. A local canvas prototype gained AI generation, node interactions and media handling.

August 2026 — Grow into a workspace. Cloud accounts, an independent asset library, an Agent, semantic search and broader multimodal creation arrived. I also began exploring subscriptions and model configuration.

September 2026 — Build a product system. Model management, Credits, subscriptions and Agent execution matured, alongside experiments with a creator identity, publishing and a creation community.

October 2026 — Reach beyond the workspace. Work continued on content governance, media processing and operations, plus the browser extension and Feishu connector.

Some features expanded while early ideas were removed or redesigned. I once explored Toolbox, Inbox and Task as separate modules, then narrowed the main product objects around canvas, Agent and assets. Its evolution has meant revisiting the information architecture, interactions and system boundaries, not simply adding features.

I tried many ways of organizing Bluetful. As modules multiplied, navigation, settings, resources and actions all became more complex. A few principles now guide the structure.

Organize around the core flow. Canvas supports visual creation, Agent understands and carries out tasks, and the asset library manages what is worth keeping. Prompts and Skills make methods reusable across the flow.

Keep automation controllable. Use different confirmation levels for different risks, avoiding repetitive prompts for low-risk work while retaining control over consequential actions.

Resist endless expansion. Not every capability needs a top-level page. The browser extension and Feishu integration, for example, can live under Connect settings.

Separate creation from lasting assets. People should be free to try, generate and revise without having to manage every temporary result forever. These decisions continue to clarify Bluetful's structure.

Bluetful began with a personal need to make an AI short film without losing materials across tools. I built a canvas, then added assets, Agent, Prompts and Skills. Moving toward SaaS also brought accounts, cloud storage, model costs, subscriptions, permissions, content moderation and reliability into view.

AI assisted development helped me turn ideas into working software more quickly and test them by using, revising and rebuilding them. Greater speed also surfaced problems sooner: feature sprawl, complex state, coupling between modules and technical debt from rapid implementation.

I have started using clearer requirements, development standards, automated tests and technical audits to manage that work. Bluetful is still in active development; some capabilities are unfinished or not publicly available, and there is not yet enough real user data to validate every assumption.

For me, Bluetful is a digital creation hub for the AI era.