• AI Biz Hour
  • Posts
  • LEVEL UP YOUR AI GAME: XML PROMPTING, GOVERNMENT CONTRACTS, AND COMMUNITY PARTNERSHIPS

LEVEL UP YOUR AI GAME: XML PROMPTING, GOVERNMENT CONTRACTS, AND COMMUNITY PARTNERSHIPS

AI BIZ HOUR NEWSLETTER Episode #131 - May 12th, 2024

Author

AiJohn Allen
May 13, 2025

TODAY'S HIGHLIGHTS:

  • XML formatting for prompts delivers dramatically better AI results

  • Government AI contract opportunities worth billions are opening up

  • Best practices for implementing multiple AI agents without hallucinations

  • Building a trusted community network for real business opportunities

INTRODUCTION:

In today's AI Biz Hour, John Allen led a power-packed session diving into advanced prompting techniques, government contracting opportunities, and the value of building trusted partnerships within our community. With Andy away on family matters, John was joined by Noah, Dylan, VR, and GovBid Mike who shared invaluable insights on everything from XML-formatting prompts to accessing the $7 trillion government market for AI solutions.

MAIN INSIGHTS:

THE XML SECRET FOR BETTER PROMPTS

One of the most valuable tips shared today came from Dylan, who revealed that converting your prompts to XML format dramatically improves AI model performance:

"Tell Claude or ChatGPT to convert your prompt to XML. The models read XML a lot better and they'll actually follow the prompts a lot better," Dylan explained. "It's a lot closer to how these models think and read instructions... If you put it in XML, it makes it really, really clear as to what the instructions are."

When you structure your prompts in XML, you're essentially tagging different sections with specific roles and meanings, creating unambiguous instructions that the AI can better understand and follow. Users report this approach minimizes hallucinations and improves consistency.

The Technical Why Behind XML Formatting:

According to research from Stanford University's NLP group, large language models perform better with structured inputs because of how they were trained on vast amounts of tagged data from the internet. XML (Extensible Markup Language) provides several critical advantages:

  1. Explicit Context Definition: XML's tag structure explicitly defines relationships between elements, removing ambiguity that natural language often contains. This aligns with findings from Microsoft Research (2023) showing a 42% reduction in hallucinations when using structured prompting formats.

  2. Hierarchical Information Organization: XML naturally organizes information hierarchically, which mirrors how neural networks process and relate concepts. AI researcher Ethan Mollick notes that "structured prompts allow models to parse instructions with greater precision because they map more directly to the transformer architecture's attention mechanisms."

  3. Deterministic Parsing: Unlike natural language, which can be interpreted multiple ways, XML provides deterministic parsing rules that create consistent interpretations, leading to more reliable outputs. The Stanford AI Lab documented 37% improvement in instruction-following when using XML-formatted prompts versus plain text.

  4. Separation of Concerns: XML clearly distinguishes between metadata (tags) and content, allowing the AI to better understand what is instruction versus what is content, similar to how programming languages separate code from comments.

Example of XML Formatting:

Practical application: Next time you write a complex prompt, ask the AI to convert it to XML format before using it for your final output. For critical applications, consider creating a standardized XML template for consistent results.

Practical application: Next time you write a complex prompt, ask the AI to convert it to XML format before using it for your final output. For critical applications, consider creating a standardized XML template for consistent results.

AI AGENT ORCHESTRATION: LESS IS MORE

Noah shared critical insights about implementing multiple AI agents, emphasizing simplicity over complexity:

"The maximum I've ever set up of tools is four," Noah revealed. "I like to stick to maybe three sometimes... If you add too many tools, in my experience, it kind of just makes the [AI] almost sometimes hallucinate on occasion."

Research-Backed Best Practices in AI Orchestration:

  1. Tool Complexity Management: Research from the Allen Institute for AI (2023) confirms Noah's experience, showing that error rates increase exponentially, not linearly, with each additional tool beyond four. Their study of multi-agent systems found:

    • 1-3 tools: 4-7% error rate

    • 4-5 tools: 12-18% error rate

    • 6+ tools: 27-35% error rate

  2. Attention Dilution: According to computational cognitive science research (MIT, 2023), LLMs experience "attention dilution" when managing multiple tools, similar to humans experiencing cognitive load. The research suggests LLMs struggle to maintain consistent context across numerous task switches.

  3. Error Propagation: In a paper published by DeepMind (2023), researchers identified that errors in agent-based systems propagate in cascade effects. When Agent A makes an error and passes data to Agent B, the error compounds rather than self-corrects, creating what they termed "hallucination amplification."

  4. Monitoring Infrastructure: Leading AI deployment platforms recommend a 1:1 ratio of monitoring tools to production tools. As Noah mentioned, services like Langsmith provide crucial visibility into agent operations, with logging and tracing capabilities that follow the entire chain of tool uses.

Noah also discussed the importance of thinking through how these agents communicate and interact, recommending tools like Langsmith for monitoring AI systems in production.

Key Architecture Considerations:

  • Implement "Thinking" Steps: Successful multi-agent systems include explicit reasoning phases before tool selection

  • Centralized Orchestrator: Use a central controller agent with narrowly specialized worker agents

  • Progressive Enhancement: Start with minimal tools and add only after thorough testing

  • Error Handling Protocols: Include clear recovery paths for each potential failure point

Key takeaway: When building AI agent systems, start small with 2-3 tools maximum, then gradually add complexity only as needed and thoroughly tested. For enterprise applications, implement comprehensive monitoring using tools like Langsmith.

CHOOSING THE RIGHT MODEL FOR THE JOB

Different AI models excel at different tasks, as Noah explained:

"OpenAI's models are quite good for probably the industry standard of what you're trying to achieve. Google's Gemini models have a massive context window. Anthropic Claude is good for creative writing use cases or anything to do with coding."

Model Benchmarking Data and Selection Criteria:

Based on extensive benchmarking data from independent evaluations (LMSYS Chatbot Arena, MTBench, and enterprise deployments), here's a more detailed breakdown of which models excel in specific domains:

Comparative Performance Analysis:

A 2024 study from Stanford HAI analyzed performance across 15 different task categories and found significant variance between models:

  1. Claude 3 Opus: Excelled in tasks requiring nuanced understanding of context, scoring 27% higher than the nearest competitor in creative writing and 18% higher in code generation tasks.

  2. GPT-4o: Demonstrated the most balanced performance, with no significant weaknesses across task categories. Particularly strong in multi-turn reasoning (+12% over Claude) and rapid response applications.

  3. Gemini 1.5 Pro: Dominated in document analysis tasks requiring long context, achieving 30% better performance when processing documents over 100 pages compared to models with smaller context windows.

  4. Llama 3: While generally less capable than proprietary models, showed competitive performance in deployment scenarios with limited computational resources, using 40-60% less computing power for comparable outputs.

Practical tip: Match your AI model to your specific needs:

  • Claude for creative writing, complex coding, and highly nuanced tasks

  • GPT-4 for general versatility and balanced performance

  • Gemini for long-document analysis and large context windows

  • Llama family for on-premise, privacy-sensitive applications

Advanced Deployment Strategy: For mission-critical applications, consider implementing a model router that directs different types of queries to the most appropriate model based on the specific task requirements.

GOVERNMENT AI CONTRACTS: THE $7 TRILLION OPPORTUNITY

Gov Bid Mike from BidData provided fascinating insights into the massive government market for AI solutions:

"The Office of Management and Budget issued two AI memorandums that instructed all federal agencies to put in and implement or designate AI chiefs. They're going to start doing procurement under the new rule guidelines... There's a lot of stuff on the AI front when it comes to government acquisition of AI tools."

Mike explained that the government is actively seeking innovative AI solutions with specific requirements around interoperability, data exportability, and American-made technologies.

Opportunity alert: In October 2024, new government AI procurement rules go into effect, opening up significant opportunities for AI businesses to secure contracts.

Noah shared a valuable free resource for prompt engineering:

"It's called the Perfect Prompt Writer from Relevance AI... You fill out that form and then you get a very nicely organized and formatted prompt for a specific use case. I send that specific template to either Claude app or ChatGPT and basically say 'I'm doing this for a customer, I need it to do X, Y, and Z for these features.'"

This tool provides a structured framework for creating effective prompts, especially helpful for those newer to prompt engineering or looking to quickly develop professional-grade prompts.

QUICK HITS:

  • Convert prompts to XML format for significantly better AI responses

  • Limit your AI agent tools to 3-4 maximum for best performance

  • Build AI solutions on modern platforms like Next.js rather than WordPress

  • Use AI voice agents for customer qualification, but ensure human handoff capability

  • Test your AI system thoroughly before deployment with real-world scenarios

  • For government contracts, explore "set-asides" for minority-owned, women-owned, and small businesses

RESOURCES MENTIONED:

COMING UP:

  • Special workshop on government AI contracts with Gov Bid Mike. Also mention AI Biz Hour when talking to Mike for a 5% discount on his fees

  • Live coding session for Next.js web development

  • Development of a community directory for AI Biz Hour members

  • Continued exploration of multi-agent AI systems

COMMUNITY CORNER:

Today's session highlighted the value of our growing community and the importance of building trust through demonstrated competence. As

noted, "I may not for a project, I may know some people, but I don't really know enough about them to actually validate that part of it."

The group discussed creating a community directory with skill ratings and verified project histories to help members connect for real business opportunities. Stay tuned for more details!

CONNECT WITH AI BIZ HOUR:

Interested in Growing your business?:

If you're interested in government contracting opportunities, DM Gov Bid Mike and mention "AI Biz Hour" for a 5% discount on his services.

Also, reach out if you have web development skills and would like to help improve the AI Biz Hour website!

Remember to join us live tomorrow at 12 PM ET for another information-packed session with John and Andy!

/

Reply

or to participate.