Flow 2 · Browser UX
Use PromptSuggestion via Chrome extension
This flow is for people who want Smart Dislike behavior directly inside a browser chat, without toggling developer mode or touching memory settings. Everything runs locally through your own MCP server.
Concept
What the extension does
Overlay a 'Smart Dislike' button
The extension injects a small Smart Dislike button near existing thumbs-down controls in your chat UI.
Grab recent messages
When clicked, it reads a small window of the last user + assistant messages from the DOM.
Call your local MCP server
It sends those messages plus your optional comment to your local MCP endpoint and waits for a single best prompt.
Insert the new prompt
The extension drops the suggested prompt into your chat input box. You can tweak it or just press Enter.
High-level setup (local MCP + Chrome extension)
- Clone and run the local MCP server (TypeScript / Claude-Desktop version):
git clone https://github.com/d180/PromptSuggestion-Claude-Desktop.git cd PromptSuggestion-Claude-Desktop npm install # Add your Gemini key echo "GEMINI_API_KEY=your_key_here" > .env # Start the MCP server npx tsx src/index.ts
By default, the server listens on http://localhost:3333/dev/analyze, which is what the extension expects.
- Install the Chrome extension:
git clone https://github.com/d180/PromptSuggestion-Chrome-Extension.git
In Chrome, go to chrome://extensions, enable Developer mode, click Load unpacked, and select the PromptSuggestion-Chrome-Extension folder.
- Open your LLM chat in the browser (for example, ChatGPT). Ask a question and wait for the assistant’s reply.
- When the answer isn’t good enough, click the 💡 Smart Dislike button on that assistant message. The extension will:
- Send the recent messages to your local MCP server
- Receive one improved, first-person prompt
- Show it in a floating panel
- Click “Use as prompt” to insert the suggestion into the chat input. You can edit it slightly if you want, then press Enter to send.
The extension never stores your conversations. It only forwards a small recent context window to your localhost MCP server so it can rewrite the prompt for you.