Prisme

See your writing through every perspective.

Prisme is an AI-powered writing tool that helps journalists and writers understand how their articles land with different audiences before they publish. Write in a rich text editor, then get instant feedback from three ideologically distinct AI agents, each powered by Mistral.

Built for the Mistral AI Hackathon by Team Bonaparte.

---

The Problem

Every article carries blind spots. A journalist may unintentionally favor one perspective, miss critical context, or use framing that alienates part of their audience. Traditional editing catches grammar not bias.

The Solution

Prisme puts three AI "lenses" on your writing simultaneously:

Each agent scores your article (1–10), flags disagreements, and can suggest targeted edits that appear as an interactive diff you can accept or reject. You can also calibrate your agents by uploading folders of articles and batch-testing them to see how each agent scores across a corpus, then tweak agent personas and descriptions until they behave exactly how you want. The entire app supports light and dark mode, and is fully localized in both English and French.

Powered by Mistral

Prisme uses Mistral Large (mistral-large-latest) exclusively for all AI features:

• Feedback generation: Each agent evaluates articles from their ideological perspective, returning structured JSON (approval score, disagreements, perspective summary)
• Improvement suggestions: Agents propose minimal, targeted edits to improve how the article reads from their viewpoint
• Batch evaluation: Run all agents against entire article folders for calibration
• JSON mode: Guaranteed structured output via Mistral's responseFormat: { type: "json_object" }

---

Running Locally

Prerequisites

• Node.js 18+ and npm
• A Mistral API key — get one at console.mistral.ai

Setup

# 1. Clone the repo
git clone https://github.com/YOUR_USERNAME/prisme.git
cd prisme

# 2. Install dependencies
npm install

# 3. Create your environment file
cp .env.example .env.local

Add your Mistral API key to .env.local:

MISTRAL_API_KEY=your_api_key_here

Run

npm run dev

Open http://localhost:3000 in your browser.

Build for production

npm run build
npm start

---

Tech Stack

Layer	Technology
Framework	Next.js 16 (App Router)
Language	TypeScript, React 19
Styling	Tailwind CSS v4
Editor	TipTap (ProseMirror)
AI	Mistral Large via @mistralai/mistralai SDK
PDF Parsing	pdf-parse
State	React Context + localStorage

Project Structure

src/
├── app/
│   ├── page.tsx              # Main editor + agent sidebar
│   ├── tune/page.tsx         # Agent tuning + batch testing
│   ├── api/
│   │   ├── feedback/route.ts # Mistral feedback endpoint
│   │   ├── improve/route.ts  # Mistral edit suggestions
│   │   └── upload/route.ts   # PDF/text file upload
│   ├── layout.tsx
│   └── globals.css
├── components/
│   ├── RichEditor.tsx        # TipTap rich text editor
│   ├── AgentCard.tsx         # Agent feedback display
│   ├── DiffView.tsx          # Interactive diff viewer
│   └── ScoreBar.tsx          # Visual score bar
└── lib/
    ├── agents.ts             # Agent definitions & personas
    ├── AgentContext.tsx       # Global state (agents, theme, locale)
    ├── i18n.ts               # EN/FR translations
    ├── diff.ts               # Word-level LCS diff algorithm
    ├── stripHtml.ts          # HTML → plain text for API calls
    └── useLocalStorage.ts    # Persistent state hook

---

Team Bonaparte

Built with care for the Mistral AI Hackathon.