跳到正文
DEV Community· Harshit Agarwal·· 3 小時前AI 評分30

Ghostwriter AI：如果 AI 故事講述者記住你犯下的每個錯誤？

Ghostwriter AI: What if the AI Storyteller Remembers Every Mistake You Make?

AI 導讀

Ghostwriter AI 是一款互動式哥德式恐怖故事生成器，會持續記錄玩家發現的線索、心理創傷、行動與抉擇，讓每個選擇在故事中留下不可逆的痕跡。它透過 Gemini API 使用 Google 開放權重 Gemma，預設模型為 gemma-4-26b-a4b-it，另提供 gemma-4-31b-it 作為高推理能力的替代選項。

正文

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 4: Ghost in the Machine

What I Built

Ghostwriter AI — Every Choice Leaves a Scar is an interactive, unscripted gothic horror storytelling generator designed to give players genuine narrative agency while confronting them with the indelible consequences of their decisions.

Traditional text adventures rely on static, pre-written decision branches: choose option A or B to jump to passage 12 or 13. Conversely, modern conversational AI chatbots often suffer from narrative amnesia, abrupt tonal drift, or cheerful assistant personas that completely dissolve suspense.

The Interpretation of "Ghost in the Machine"

In gothic folklore, a haunting is an unresolved echo of past trauma trapped inside an environment. In Ghostwriter AI, the "Ghost" is the persistent memory ledger trapped within the machine.

The machine does not just generate words—it holds a grudge. As you explore eerie domains (from the decaying halls of Blackwood Manor to abandoned sanitariums and cursed hamlets), the engine continuously tracks your discovered clues, psychological wounds, unprompted actions, and fateful decisions. Every choice leaves an irreversible scar on the narrative ledger, culminating in a climactic autopsy ending (Escape, Survival with a cost, Tragic ending, Supernatural revelation, or Hidden ending).

Demo

Visual Preview & Experience

The application features a dark gothic aesthetic inspired by classic horror literature:

  • Atmospheric Palette: Near-black charcoal backgrounds (#0a0a0c), dynamic vignette, subtle SVG noise film grain, and restrained blood-red accents.
  • Cinematic Typography: Cinzel for regal titles and headers, EB Garamond with dedicated dialogue styling for story passages, and Inter for accessible UI controls.
  • Interactive Consequence Ledger: The collapsible "What the darkness remembers" panel dynamically displays clues, inventory relics, character states, and permanent scars.
  • Unscripted Agency: Players can choose from three dynamically generated choices (with keyboard shortcuts [1], [2], [3]) or type freeform custom actions (e.g., "Press your fingertips against the peeling wallpaper near the hearth to feel for drafts").

Code

Ghostwriter AI — Every Choice Leaves a Scar

An unscripted interactive horror-story generator where the machine remembers your decisions, powered by Google's open-weight Gemma model via the Gemini API.
Built for the DEV.to Hacktoberfest Week 4 2026 Challenge: "Ghost in the Machine".

Ghostwriter AI Banner

📖 The Problem and Creative Concept

Traditional text adventures and branching narrative games rely on pre-written decision trees. If you choose Door A, you get pre-baked Paragraph 12; if you choose Door B, you get pre-baked Paragraph 13. While predictable, this approach limits player agency and quickly loses tension once branches are discovered.

On the other hand, unstructured conversational AI chatbots lack memory discipline: they frequently suffer from tone drift, forget past decisions within a few turns, and lack pacing or stakes.

Ghostwriter AI bridges this gap It acts as an intelligent, literary storyteller that weaves genuine gothic horror. Every choice you select—or custom unscripted action you…

Direct Repository Link: https://github.com/Harshitagarwal113/GhostWriter-AI

How I Built It

Ghostwriter AI is built as a decoupled, high-performance full-stack web application designed for deterministic state management and zero-leak AI inference.

GhostWriter AI/
├── frontend/                 # React 19 + TypeScript + Tailwind CSS + Vite
│   ├── src/
│   │   ├── components/       # AtmosphericBackground, StoryChoices, DarknessLedger, etc.
│   │   ├── pages/            # LandingPage, StorySetupPage, StoryPlayerPage, EndingPage
│   │   ├── hooks/            # useStorySession, useApiHealth
│   │   ├── services/api.ts   # Centralized Fetch API client with timeouts & retry logic
│   │   └── types/story.ts    # Type-safe TypeScript contracts
├── backend/                  # Python 3.14 + FastAPI + Pydantic v2 + google-genai
│   ├── app/
│   │   ├── api/routes.py     # /api/health, /api/story/start, /api/story/advance
│   │   ├── services/         # gemma_client.py, story_engine.py, story_memory.py
│   │   ├── prompts/          # storyteller.py (gothic system instructions & templates)
│   │   └── main.py           # FastAPI server with CORS & error handlers
│   └── tests/                # 13 automated Pytest unit and integration tests
└── README.md

The AI Engine: Google's Open-Weight Gemma Model

  • Model: Google's open-weight Gemma model (gemma-4-26b-a4b-it default, gemma-4-31b-it high-reasoning alternative).
  • Integration: Implemented through Google's official Gen AI Python SDK (google-genai) and the Gemini API delivery endpoint.
  • Bounded Retry & JSON Sanitation: Language models can occasionally emit markdown code fences or non-JSON preamble. Our gemma_client.py strips fences, isolates JSON object boundaries, and executes a bounded single-retry fallback to guarantee valid structured schemas without redundant token consumption.
  • Pacing Engine: To prevent open-ended AI rambling, story length budgets (Short: 3–4 scenes, Medium: 5–7 scenes, Long: 8–10 scenes) guide Gemma from escalating tension directly into a fateful climax.
  • Automated Test Suite: 13 automated unit tests cover health reporting, Pydantic input validation, mocked story progression, custom action resolution, and CORS preflight.

Why Does Open Innovation Matter?

Open innovation and open-weight models like Gemma are transformative for interactive fiction and creative AI development:

  1. Transparency & Auditable Alignment: Unlike proprietary black-box APIs whose system prompts, tokenizers, and alignment penalties can change without notice, open-weight models give developers full insight into model behavior, allowing creators to craft precise narrative systems with confidence.
  2. Creative Literary Expression Without Sanitization: Horror fiction relies on psychological tension, dread, and eerie themes. Closed commercial models often over-censor atmospheric horror prompts into bland cheerful statements. Open models respect literary nuance while keeping safety boundaries intact.
  3. Reproducibility & Longevity: An interactive story built on open standards won't break if a vendor deprecates a proprietary endpoint. Developers can run open weights locally, fine-tune them on gothic literature corpora, or deploy them through hosted APIs.

My Agent Session

This application was developed, tested, and integrated with the assistance of agentic pair-programming workflows:

Conversation ID: acff272d-74b8-48dc-9bf7-171b332d47c4
Tooling: DevRelay Gateway, Google Gen AI Python SDK, Browser Subagent Verification

The Spirit of It

During our live test of the Haunted Mansion scenario, the protagonist Eleanor encountered a weeping portrait over an ancient hearth. Instead of taking a pre-generated choice, we tested the custom action input:

Player Action: "Press your fingertips against the peeling wallpaper near the hearth to feel for drafts."

Gemma didn't give a canned response. It adapted to her exact touch:

"Eleanor's fingers brush against the desiccated, floral patterns of the wallpaper... beneath the layer of decay, her touch finds the unmistakable, sharp edge of a recessed latch, cold as a tombstone. A hollow, rhythmic tremor shudders through the wall—vibrating like a biological pulse."

When she manipulated the latch in the final scene, the climax landed with true chills:

"There is a wet, sickening sound—the sound of a suture tearing through muscle. The wall does not swing on hinges; it peels back like a heavy, weeping eyelid... Eleanor discovers that Blackwood Manor is not a house, but a predatory, sentient organism."

The memory ledger immediately updated her status to "Sensory assimilation" and "Fatal entrapment", sealing her fate in "The Eternal Pulse" ending. That was the moment the project truly came alive: the machine wasn't just predicting the next token—it was remembering her curiosity and turning it into her undoing.

Prize Categories

  • Hacktoberfest Open-Source AI Challenge Week 4: "Ghost in the Machine"

來源：DEV Community · dev.to

#開源/倉庫#開源生態#Google