Listen to this Post
Introduction:
The modern knowledge worker’s day is a symphony of context switches: Google Gemini for editing, Grok for real-time information, Microsoft Copilot for work, Anthropic Claude for thoughtful answers, and OpenAI ChatGPT for direct responses. Remembering which LLM responds best to which prompt structure is an exercise in frustration—and LLMs are “too nice” to tell you what’s missing from your prompt until you’ve already wasted cycles going down a rabbit hole. Chat Copilot, a Chrome extension built by Lighty AI CEO Richard Rabbat (former Twitter VP of Product), addresses this exact pain point by reviewing your prompts before you send them, making suggestions, and tracking your “efficiency score”—all powered by Gemini Nano running locally in your browser.
Learning Objectives:
- Understand the concept of prompt fatigue and why context switching between multiple LLM interfaces degrades productivity more than prompt quality itself
- Learn how Chat Copilot leverages Chrome’s built-in Prompt API and Gemini Nano to provide on-device, privacy-first prompt optimization
- Master the five prompt quality metrics (clarity, specificity, structure, completeness, actionability) used to score and improve your LLM interactions
- Gain hands-on knowledge of enabling Chrome’s built-in AI capabilities and configuring the extension for optimal performance
- Explore prompt engineering best practices and how to systematically improve your prompts across different LLM platforms
You Should Know:
- The Prompt Fatigue Problem – Why Context Switching Kills Productivity
The real issue isn’t prompt quality—it’s prompt fatigue. Most people think they need better prompts when they actually need to stop context switching between six different interfaces. Every prompt is a context switch. Every review is a recovery from a context switch. The fix isn’t to prompt better; it’s to structure work so the AI can run without you constantly in the loop.
Chat Copilot was built with this specific use case in mind. The extension serves as your copilot for LLMs, improving your prompt based on which LLM you’re using and adapting suggestions based on each model’s capabilities. It supports nine platforms: ChatGPT, Claude, Gemini, Grok, and Chinese models like 通义千问, 文心一言, 腾讯元宝, and DeepSeek.
Step‑by‑step: Installing and Configuring Chat Copilot
- Install from Chrome Web Store: Navigate to the Chat Copilot listing (extension ID: pgjafgigjhkpboaklnlgcidhpidikakl) and click “Add to Chrome”
- Pin the extension: Click the puzzle piece icon in your Chrome toolbar and pin Chat Copilot for easy access
- Navigate to your preferred LLM: Open ChatGPT, Claude, Gemini, Copilot, or Grok in your browser
4. Write your prompt as you normally would
- Click the Chat Copilot icon to review and improve your prompt before sending
- Review your efficiency score: The extension scores your prompt across five metrics—clarity, specificity, structure, completeness, and actionability
- Apply suggestions and re-score until you achieve a high-quality prompt
-
Under the Hood – Chrome’s Prompt API and Gemini Nano
Chat Copilot leverages Chrome’s built-in Prompt API, which exposes Gemini Nano through the `LanguageModel` namespace—available on `window.LanguageModel` for web pages and `chrome.languageModel` for extensions. The model runs entirely on-device with no network calls. While the API is built into Chrome, the model is downloaded separately the first time an extension uses the API.
Step‑by‑step: Enabling Chrome’s Built-in AI (Required for Chat Copilot)
Before Chat Copilot can use Gemini Nano, you must enable Chrome’s built-in AI features:
- Open Chrome flags: Navigate to `chrome://flags` in your browser address bar
- Enable Prompt API: Search for `prompt-api-for-gemini-1ano` and set it to “Enabled”
- Enable Optimization Guide: Search for `optimization-guide-on-device-model` and set it to “Enabled BypassPerfRequirement”
- Restart Chrome: Click the “Relaunch” button that appears at the bottom of the flags page
- Verify model download: Navigate to
chrome://components/, find “Optimization Guide On Device Model,” and ensure it’s fully downloaded - Alternative verification: Check `chrome://on-device-internals` to confirm Gemini Nano is available
Linux/macOS Command (for troubleshooting):
Check Chrome version (must be 127+ for Prompt API support) google-chrome --version or on macOS /Applications/Google\ Chrome.app/Contents/MacOS/Google\ Chrome --version
Windows Command (for troubleshooting):
Check Chrome version wmic datafile where name="C:\Program Files\Google\Chrome\Application\chrome.exe" get Version Alternative reg query "HKEY_CURRENT_USER\Software\Google\Chrome\BLBeacon" /v version
- The Five Metrics – What Makes a “Good” Prompt?
Chat Copilot scores your prompts across five dimensions:
- Clarity: How clear and unambiguous your instructions are. Higher clarity means the AI understands exactly what you want
- Specificity: Level of detail and precision in your requirements. More specific prompts generate more targeted responses
- Structure: How well-organized your prompt is—logical flow, proper formatting, and clear sections
- Completeness: Whether you’ve included all necessary context, background, assumptions, and domain information
- Actionability: Whether your prompt leads to concrete, executable outputs rather than vague or open-ended responses
Step‑by‑step: Improving Your Prompt Score
- Start with a rough draft: Type your initial prompt as you normally would
- Run Chat Copilot: Click the extension icon to analyze your prompt
- Review the score breakdown: Identify which metric scored lowest
4. Apply specific improvements:
- For low clarity: Remove ambiguous language; specify exactly what you want
- For low specificity: Add numbers, scope, constraints, or examples
- For low structure: Use bullet points, numbered lists, or clear section headers
- For low completeness: Add background context, assumptions, or relevant domain knowledge
- For low actionability: Request specific output formats (JSON, table, code, etc.)
- Re-run the analysis to see your improved score
6. Send your optimized prompt to the LLM
4. Privacy Architecture – Why Local Processing Matters
Chat Copilot’s privacy model is built around a core principle: all prompt processing happens locally using Gemini Nano running in your browser. The extension does not send your prompts to external servers. According to the developer’s privacy policy, user data is not sold to third parties, not used or transferred for purposes unrelated to the product’s core functionality, and not used to determine creditworthiness or for lending purposes.
Step‑by‑step: Verifying Privacy and Local Processing
- Open Chrome DevTools: Press `F12` or `Ctrl+Shift+I` (Windows/Linux) / `Cmd+Option+I` (macOS)
- Go to the Network tab: Monitor all network requests while using Chat Copilot
- Use the extension: Write a prompt and click the Chat Copilot icon
- Verify no external requests: You should see no network calls to external AI APIs—all processing stays local
- Check extension permissions: Navigate to
chrome://extensions/, find Chat Copilot, and click “Details” to review permissions
Code Snippet: Checking if Prompt API is Available (for Developers)
// Check if the Prompt API is available in Chrome
async function checkPromptAPI() {
if (!window.ai || !window.ai.languageModel) {
console.log('Prompt API not available. Enable chrome://flags/prompt-api-for-gemini-1ano');
return false;
}
const availability = await window.ai.languageModel.availability();
console.log('Model availability:', availability);
// Returns: 'readily', 'after-download', or 'no'
return availability;
}
checkPromptAPI();
- Prompt Engineering Best Practices – Beyond the Score
While Chat Copilot helps optimize your prompts, understanding the underlying principles of prompt engineering will make you a more effective LLM user overall. Research shows that prompt engineering is the practice of designing, versioning, testing, and optimizing the text, structure, and tool interfaces that condition an LLM toward a target behavior.
Step‑by‑step: Systematic Prompt Optimization
- Start with clarity: Be specific about what you need. Remove ambiguity and state your goal explicitly
- Add context: Context helps AI generate accurate responses that best fit a specific scenario. A prompt that lacks context becomes generic and gives room for ambiguity
- Use structure: Organize your prompt with clear sections, bullet points, or numbered lists
- Specify output format: Tell the LLM exactly how you want the response formatted (JSON, table, markdown, code, etc.)
- Include examples: Few-shot examples help the LLM understand the expected output pattern
- Iterate and refine: Use Chat Copilot’s scoring to identify weak points and continuously improve
Example: Before and After Prompt Optimization
Before (Score: ~40/100):
> “Write a blog post about AI.”
After (Score: ~85/100):
“Write a 1,200-word technical blog post for a CTO audience about implementing on-device AI using Chrome’s Gemini Nano. Include: (1) an overview of the Prompt API, (2) step-by-step implementation guide with code examples, (3) privacy and security considerations, and (4) performance benchmarks. Format with clear H2 headings and include actionable takeaways. Tone: professional but accessible.”
6. Troubleshooting and Advanced Configuration
Common Issues and Solutions
| Issue | Solution |
|-|-|
| “Prompt API not available” | Ensure Chrome version 127+ (Canary/Dev recommended) |
| Model not downloading | Check `chrome://components/` and click “Check for Update” on Optimization Guide On Device Model |
| Extension not appearing | Pin the extension from Chrome’s puzzle piece icon |
| Low efficiency scores | Review the five metrics and systematically address each weakness |
| Privacy concerns | Verify no network requests in DevTools Network tab |
Advanced Configuration
For developers who want to customize the experience, Chat Copilot can be built from source:
git clone https://github.com/hellolib/chat-copilot.git cd chat-copilot npm install && npm run build Load the extension from build/chat-copilot in chrome://extensions/ with Developer Mode enabled
Windows PowerShell Command (for checking Chrome version):
Get-ItemProperty "HKLM:\SOFTWARE\WOW6432Node\Microsoft\Windows\CurrentVersion\Uninstall\Google Chrome" | Select-Object DisplayVersion
Linux Command (for checking if model is downloaded):
Check the model files in Chrome's profile directory ls -la ~/.config/google-chrome/OptimizationGuideOnDeviceModelStore/ Look for files related to Gemini Nano (approximately 4GB in size) du -sh ~/.config/google-chrome/OptimizationGuideOnDeviceModelStore/
What Undercode Say:
- Prompt fatigue is the real enemy, not prompt quality. Most users don’t need “better” prompts—they need to stop context switching between six different interfaces and develop a systematic approach to LLM interaction. Chat Copilot addresses this by providing real-time feedback and scoring that trains users to write better prompts organically.
-
Privacy-first local AI is the future of productivity tools. By leveraging Chrome’s built-in Gemini Nano, Chat Copilot demonstrates that powerful AI assistance doesn’t require sending sensitive data to the cloud. This architecture sets a new standard for enterprise-ready AI tools that respect data sovereignty.
-
Metrics change behavior. Tracking efficiency scores across five dimensions (clarity, specificity, structure, completeness, actionability) provides actionable feedback that actually changes how people write. Unlike vague “good prompt” advice, scored metrics create a clear improvement path.
-
The extension model is evolving. As Chrome’s built-in AI capabilities mature (stable for extensions in Chrome 138+), we’ll see more sophisticated on-device AI tools that rival cloud-based offerings. Chat Copilot is an early indicator of this shift.
-
Adoption barriers remain. Users must enable experimental Chrome flags and download a ~4GB model before using the extension. This friction point will need to be addressed as built-in AI becomes the default in stable Chrome releases.
Prediction:
-
+1 On-device AI will become the default for privacy-sensitive enterprise workloads within 18–24 months, reducing reliance on cloud APIs for routine LLM interactions and cutting latency by 40–60%.
-
+1 Prompt optimization tools like Chat Copilot will evolve into full-featured “prompt IDEs” that include version control, A/B testing, and team collaboration features—democratizing prompt engineering expertise.
-
-1 The fragmentation of LLM interfaces will worsen before it improves, as more specialized models enter the market. Users will need even more sophisticated orchestration layers to manage their AI workflows effectively.
-
+1 Chrome’s built-in AI will accelerate the development of a new class of browser-1ative AI applications that work offline, with no API costs, and complete data privacy—fundamentally changing how we think about AI-powered browsing.
-
-1 The ~4GB model download requirement will remain a barrier for casual users, potentially limiting adoption to power users and developers until Chrome ships with the model pre-installed.
About the Developer: Chat Copilot is developed by Lighty AI, Inc., founded by Richard Rabbat (former VP of Product at Twitter) and Anthony Johnson (former CTO of Giphy). The company is focused on AI agents that help users navigate their workday more efficiently.
▶️ Related Video (62% Match):
🎯Let’s Practice For Free:
🎓 Live Courses & Certifications:
Join Undercode Academy for Verified Certifications
🚀 Request a Custom Project:
Secure, high-velocity infrastructure and disruptive technological engineering. Contact our engineering team for high-tier development and proprietary systems:
[email protected]
💎 Smart Architecture | 🛡️ Secure by Design | ⭐ Trusted by Thousands
IT/Security Reporter URL:
Reported By: Richard Rabbat – Hackers Feeds
Extra Hub: Undercode MoN
Basic Verification: Pass ✅
