Novel Cognitive Engine: Resilient AI Writing Workspace

• 

  NCE keeps the writing session alive by routing around blocked AI generation paths.

• 

  The Review Room shows the exact fallback path, local request log, and Firestore mirror status.

• 

  When OpenAI is unavailable and Gemini is disabled, NCE falls through to local_worker.

• GIF

  The fallback demo preserves draft progress instead of failing the author session.

Inspiration

Most AI writing tools are good at producing a page of text, but they become unreliable when the project turns into a long-form manuscript. Characters forget what they know, clues leak too early, revisions overwrite canon, and planning gets disconnected from drafting.

We wanted to build a writing system that treats long-form fiction more like a managed production workflow than a one-shot prompt. That became Novel Cognitive Engine, or NCE Workspace.

What it does

NCE Workspace is a continuity-aware writing workspace for long-form fiction.

It gives an author one connected workflow:

1. Open or create a novel workspace
2. Inspect planning artifacts in Design Room
3. Draft and revise chapters in Writing Room
4. Run validation in Review Room
5. Save only when the chapter passes story-state and quality gates

In the current demo, the product supports:

• prewrite artifacts
• story bible editing
• narrative state inspection
• design preflight
• context preview
• scene contract review
• thread review
• narrative debug reports
• continuity-aware save gating
• valid save / invalid save rejection
• provider-priority routing with fallback
• local and Firestore-mirrored generation logs

How we built it

We built NCE as a writing workspace with explicit workflow boundaries instead of a monolithic text generator.

• Next.js + React for the workspace UI
• Python + FastAPI for the backend
• MCP tools for orchestration
• persisted workspace state for planning artifacts, story bible, and chapter flow
• validation layers for continuity, quality, and save safety
• provider-priority routing for generation fallback
• local request persistence with optional Firestore mirroring

In the current demo, the generation path can skip unavailable or disabled providers, fall through to a lower-priority route, generate the next chapter draft, and preserve the request state.

Challenges we ran into

• long-form fiction quality is not just prose quality; it depends on state consistency across chapters
• save-time validation needed to block bad drafts without breaking normal author flow
• provider fallback had to degrade gracefully while still preserving request logs and user-visible state
• hackathon demo truth had to stay honest: we did not want to present future SaaS plans as already shipped

Accomplishments that we're proud of

• built a real workspace loop instead of a generic “write me a chapter” demo
• connected planning, drafting, validation, and save safety in one product surface
• implemented continuity-aware invalid-save rejection
• added provider-priority routing with local fallback and request logging
• mirrored generation requests to Firestore
• prepared a real fallback demo showing that the writing session continues even when the primary AI route is blocked

What we learned

• long-form fiction fails when planning, state, and revision are separated
• AI writing products need deterministic save-time guardrails, not just better prompts
• resilience is also a UX problem: if generation degrades, the author still needs a coherent workflow
• the strongest demo is a truthful one; clearly separating implemented behavior from future architecture improves trust

What's next

• broader provider adapter coverage beyond the current runtime
• queue persistence beyond request-log storage
• multi-user product layer and cloud storage
• human-assisted recovery workflow for failed generation runs
• deeper internationalization beyond the current bilingual demo UI

Built With

• fastapi
• firebase-firestore
• mcp
• next.js
• playwright
• python
• react
• recordly
• typescript