No API keys. No accounts. No data leaves your machine.
PromptForge is a chatbot first, prompt generator in parallel. It runs entirely on your local machine via Ollama — open a conversation, explore ideas freely, then fire a single slash command to convert the whole thread into a production-quality prompt for any AI tool.
The insight it's built on: you already know what you want — you just need help saying it right.
You say: "a girl standing in a garden"
──── /image ────▶
PromptForge: A young woman with flowing auburn hair stands in a sun-drenched
English cottage garden at golden hour, wearing a white linen dress.
Soft bokeh roses in the foreground. Cinematic lighting, 50mm lens,
shallow depth of field, photorealistic, 8K UHD. Shot by Annie Leibovitz.
Negative: cartoon, anime, blur, watermark, lowres, flat lighting
Best for: Midjourney · Flux · Stable Diffusion
┌─────────────────────────────────────────────────────────────────────┐
│ 1. Chat freely — no setup needed │
│ │
│ "I want a login form in React" │
│ "TypeScript, with JWT and CSRF protection" │
│ "Should pass WCAG AA accessibility" │
│ ↕ back & forth │
│ 2. Type a slash command when ready │
│ │
│ /code /image /video /project /brainstorm /prd │
│ │
│ 3. PromptForge reads the full conversation and synthesizes │
│ a production-quality prompt — not just your last message │
│ │
│ The thread stays intact. Chat and prompts live together. │
└─────────────────────────────────────────────────────────────────────┘
Land in a chat interface with no mode selected, no forced setup. Talk freely — ask questions, think out loud, explore an idea. The AI is a fully capable local assistant powered by whichever Ollama model you have. Prompt generation is an option, never an obligation.
Type /brainstorm to enter a focused idea-exploration session. The assistant shifts into a dedicated mode: it asks one clarifying question per turn, surfaces constraints you haven't considered, and keeps responses tight (under 120 words) to maintain a productive back-and-forth rhythm. When you're ready, one slash command converts the whole conversation into a prompt.
Every mode uses a domain-specific system prompt engineered to produce genuinely useful output, not generic rewrites:
| Mode | Slash | What gets added to your idea |
|---|---|---|
| 🖼️ Image gen | /image |
Subject, lighting, lens spec, art style, color palette, mood, negative prompt |
| 🎬 Video gen | /video |
Camera movement, shot type, duration, frame rate, scene transitions, style ref |
| 💻 Coding | /code |
Language, architecture, validation, security, tests, edge cases, accessibility |
| 📋 Project / PRD | /project |
Goals, personas, MoSCoW features, metrics, timeline, risks, out-of-scope |
| 📋 Full PRD | /prd |
Complete Product Requirements Document — sections, OKRs, user flows, tech spec |
| 🤖 AI Agent | /agent |
Persona, capabilities, constraints, tool access, failure modes, few-shot examples |
| ✍️ Creative | /creative |
POV, tense, tone, pacing, sensory details, conflict arc, character depth |
| 🔍 Research | /research |
Scope, methodology, output format, citation style, sub-questions, counter-arguments |
| 🗣️ Chat Prompt | /chat |
Context, desired format, tone, audience, examples, constraints |
Tell PromptForge which tool you're targeting — by slash command or plain English — and it formats the output to match that platform's exact conventions:
/claude /midjourney /cursor /sora
/stable-diffusion /flux /opencode /runway
/nanobanana /copilot /windsurf /comfyui
... and 13 more
Or just say "give me a prompt for Cursor" — the app parses your intent automatically.
Attach files and PromptForge injects their content as context for enhancement:
| Type | What's extracted |
|---|---|
| Full text | |
| Images | Description via vision model |
Code files (.py .js .ts) |
Structure and function signatures |
| CSV / JSON | Schema and sample data |
| Markdown / TXT / DOCX | Full content |
Every enhanced prompt is auto-saved. Browse examples, filter by mode or platform, copy with one click, or load any saved prompt back into the chat to continue from where you left off.
Type / anywhere in the input to open a live-filtered command palette. Every mode, platform, and special command is one keystroke away:
/brainstorm Explore ideas collaboratively before committing to a prompt
/image Image generation prompt
/video Video generation prompt
/code Production-ready coding prompt
/project Project brief / PRD prompt
/prd Full Product Requirements Document
/agent AI agent system prompt
/creative Creative writing prompt
/research Structured research prompt
/chat Conversational prompt clarity
/search Web search with cited sources
/prompt Prompt it — enhance the current conversation now
/claude Target Claude · /midjourney · /cursor · /opencode ...
Every conversation auto-saves to a local .md file. Restart the app and resume exactly where you left off — message history, active mode, platform, and your notes, all preserved.
- Ollama installed and running
- Python 3.10+
- Node.js 18+
# Primary reasoning model — pick based on available VRAM
ollama pull hermes3:8b # 8 GB — excellent for prompt engineering
ollama pull qwen3:4b # 4 GB — solid fallback
# Coding specialist (recommended for /code and /prd modes)
ollama pull qwen2.5-coder:7b
# Embedding model (for semantic library search)
ollama pull nomic-embed-textcd backend
pip install -r requirements.txt
uvicorn main:app --reload --port 8000Windows:
py -m pip install -r requirements.txt·py -m uvicorn main:app --reload --port 8000
cd frontend
npm install
npm run devOpen http://localhost:5173 — start chatting.
You: I want a cyberpunk cityscape for Midjourney
PromptForge: Neon-soaked rain streets or dry brutalist megacity?
You: Neon rain, Blade Runner feel, lone figure silhouette
PromptForge: Wide shot or closer — should the city feel huge around them?
You: Wide shot, city dwarfs them
You type: /image
PromptForge ✦ A lone silhouetted figure stands on a rain-slicked neon-lit
street at the base of towering brutalist skyscrapers. Glowing
holographic advertisements reflect in deep puddles. Cinematic
wide shot, anamorphic lens flare, Blade Runner 2049 palette
(deep teal + electric orange), 8K, sharp focus, film grain.
Negative: flat lighting, daytime, cartoon, blur, watermark
Best for: Midjourney · Flux · Stable Diffusion
Type: /code build a JWT login form in React with TypeScript
PromptForge ✦ Create a production-ready React login form using TypeScript
and React Hook Form. Requirements: email + password fields
with Zod validation, WCAG 2.1 AA accessible labels and error
messages, show/hide password toggle, loading state on submit,
JWT stored in httpOnly cookie (no localStorage), CSRF token
header support. Handle: 429 rate limiting, network errors,
expired token redirect. Include React Testing Library tests.
Use TailwindCSS for styling.
Type: give me a prompt for Cursor to build a REST API
→ Detects "Cursor" → activates code mode
→ Asks clarifying questions about the API
→ "✦ Prompt it" generates output in Cursor's preferred format
| Action | Shortcut |
|---|---|
| Send message | Enter |
| Enhance conversation | Ctrl+Enter |
| Open command palette | / |
| Switch to image mode | /image |
| Switch to code mode | /code |
| Start brainstorm | /brainstorm |
| Generate full PRD | /prd |
promptforge/
├── backend/
│ ├── main.py # FastAPI app, CORS, router registration
│ ├── routers/
│ │ ├── enhance.py # POST /enhance, POST /enhance/chat
│ │ ├── library.py # Prompt library CRUD
│ │ ├── files.py # File upload + text extraction
│ │ └── models.py # Ollama model list
│ ├── services/
│ │ └── meta_prompts.py # All system prompts per mode
│ └── requirements.txt
│
└── frontend/
└── src/
├── App.jsx # Root — session state, enhance/chat logic
├── components/
│ ├── ChatArea.jsx # Message feed, enhanced prompt cards
│ ├── InputBar.jsx # Textarea, slash commands, file attach
│ ├── Sidebar.jsx # Mode nav, Ollama status
│ ├── Topbar.jsx # Model picker, Library toggle
│ ├── PromptLibrary.jsx # Saved + example prompt browser
│ ├── CopyButton.jsx
│ └── PlatformTags.jsx
├── lib/
│ ├── ollama.js # Ollama API client, model picking, JSON parsing
│ └── providerIntent.js # "give me a prompt for X" detection
├── hooks/
│ └── useModels.js # Model list + per-mode auto-selection
└── constants/
├── modes.js # Mode definitions, meta-prompts, slash commands
├── platforms.js # 25 supported AI platforms
└── platformGuides.js # Per-platform formatting guides
PromptForge knows the formatting conventions for 25 AI tools:
| Category | Platforms |
|---|---|
| Chat AI | ChatGPT, Claude, Gemini, Kimi, Perplexity, Grok, Meta AI, Ollama |
| Image | Midjourney, DALL·E, Stable Diffusion, Flux, Nanobanana, Ideogram, Leonardo AI, ComfyUI |
| Video | Sora, Runway, Kling, Pika |
| Code | Cursor, GitHub Copilot, Antigravity, Windsurf, OpenCode |
Everything stays on your machine:
- Ollama runs models locally at
localhost:11434— no cloud calls - No telemetry, no accounts, no API keys required
- Uploaded files are processed in memory, never stored on disk
- Prompt library is in a local SQLite file (
backend/promptforge.db) - Session context saved to a local
.mdfile you own and control
| Layer | Technology |
|---|---|
| Frontend | React 18 + Vite + TailwindCSS |
| Backend | FastAPI (Python 3.10+) |
| LLM runtime | Ollama (local, any model) |
| Storage | SQLite for prompts · .md file for session context |
| Dev proxy | Vite → Ollama (/ollama) + FastAPI (/api) |
- MCP server — expose PromptForge as a tool inside Claude Desktop and other MCP clients
- Browser extension — "Enhance with PromptForge" button injected into ChatGPT, Claude, Midjourney, and Cursor
Built for people who know what they want but need help saying it right.