Hey folks! Captain YAR here, and you won’t believe what I stumbled upon. Ever waded through a dense government contract and thought, "There’s gotta be a better way to figure this out"? Well, it looks like a clever staffer at the VA (Veterans Affairs), part of DOGE (Department of Government Efficiency, how cool is that name?!), is doing just that by unleashing AI on ’em!
✨ What’s the Big Deal?
They’re using a Large Language Model (LLM) to analyze the first 10,000 characters of government contracts and decide if they’re "munchable." Yep, "munchable"! That’s their awesome term for figuring out if a contract can be cut, its services brought in-house, or if it’s absolutely vital. The kicker? They’re using an LLM that doesn’t even have specific training in government contracts! It’s all in the prompt, baby!
⚙️ The AI’s marching orders – System Prompt Highlights:
This is how they’re priming the AI. It’s told to act like a super-smart government contract analyst. Its main gigs are to:
- 📌 Spill the Beans: Clearly explain WHO the contract is with, WHAT specific services/products are involved, and WHO ultimately benefits.
- 🚫 Hands Off! (NOT Munchable): If it’s for critical stuff like EMR systems or healthcare IT that directly supports patient care, it’s a no-go for munching.
- ✅ Fair Game! (MUNCHABLE): Contracts for things like DEI (Diversity, Equity, and Inclusion) initiatives or services that could easily be handled by good ol’ W2 employees get the "MUNCHABLE" tag. Also, many ‘soft services’ like healthcare technology management, data management, and administrative consulting are considered munchable.
- 🤷 N/A Status: Contract modifications? Those get marked as ‘N/A’ for munchability.
- 🔪 IDIQ Contracts: These (Indefinite Delivery, Indefinite Quantity) face more scrutiny for termination unless they’re for core medical services or benefits processing.
📜 The Official "Munchable" Rulebook – Deeper Dive:
This part gets into the real nitty-gritty of how the AI should make its decisions. It’s packed with details!
- Patient Care is King (NOT MUNCHABLE): Anything directly involving patient care, think physical exams, giving meds, medical procedures, is off-limits for munching.
- Can We Do It Ourselves? (MUNCHABLE): If services can be insourced (like video production, basic IT support, admin tasks), they’re likely munchable.
- Lots of Audits are Safe (NOT MUNCHABLE): There’s a big list of important exceptions, especially for audits and compliance. Things like third-party financial audits, medical equipment certifications, and healthcare facility accreditation reviews are NOT munchable.
- Other Key Rules:
- 💡 Services multiple steps removed from actual veteran care? Likely MUNCHABLE.
- 💡 Drug lifecycle management (like pricing analysis, not the actual drugs)? MUNCHABLE.
- 💡 Backup administrative roles? MUNCHABLE (seen as potentially duplicative).
🚀 Why This is a Game-Changer!
I think this is just fantastic! It shows how a well-crafted prompt can turn a general AI into a specialized analysis tool. Imagine the potential for efficiency and focusing resources where they’re needed most. This isn’t just about saving cash; it’s about making government work smarter for our veterans. This whole setup is a masterclass in practical prompt engineering!
This is a seriously detailed approach to guiding an AI. If you want to see the exact wording of the prompts and all the specific rules they’ve laid out, you’ve gotta check out the full Reddit post. It’s a goldmine for anyone interested in real-world AI applications!
Prompt used by DOGE @ VA for contract analysis
byu/STGItsMe in