How We Taught Claude to Analyze any Federal Agency's Procurement Behavior

The first time took a while and some back and forth. The second time took two prompts. By the third, we did it live on camera just to prove it works.

Here's what happened, and why it matters if you work in government contracting.

The Problem Every BD Professional Knows

If you're a government contracting professional, chances are you've done something like this before. You've got an opportunity in your inbox and you're trying to figure out: how does this agency actually buy things?

Does it use best value? Lowest price technically acceptable? Does it have preferred contract types — firm fixed price or time and materials? Does it tend toward particular set-asides? Does it use specific vehicles?

All of these questions are answerable using data. You can dig through SAM.gov, pull historical FPDS data, look at USASpending, read old solicitations. You probably end up dumping it all into Excel anyway. It works, but it takes hours, and the next time someone asks about a different agency, you start over.

We wanted to see if we could use Claude and Tango's API to do this with more automation. And then we wanted to see if we could make that automation repeatable for any agency.

The First Attempt: Forest Service

We started by asking Claude a straightforward question: How often does the Forest Service use best value or lowest price procurement for professional services?

Now, if you asked Claude this question normally, it would get some pretty weird results, because it's not actually looking at underlying contracting data. But with Tango connected to Claude via the Tango MCP, you're not just asking the internet; you're querying actual award data and putting it into the context window.

So Claude started working. It looked for the available tools, figured out the NAICS and PSC codes for professional services, pulled the contracts, and started to dive in.

But even with Tango, Claude didn't do a great job out of the box.

It didn't look at trade-off versus lowest price technically acceptable. So we called that out; we told it to look at the solicitation procedures field for contract awards. It did more work, but Claude didn't realize that solicitation procedures are not the same thing as evaluation methods. They're different fields entirely. And Claude didn't even realize it had the ability to pull the trade-off data through Tango.

Now, in fairness to Claude, govcon research can be challenging even for more-advanced practitioners. Expecting an LLM to "one-shot" this research would be pretty unrealistic.

But, once we got that sorted, things started to click. We asked Claude to look into contract types. We asked Claude about whether contract sizes might affect procurement methods. We asked about protests, whether the digests suggest anything about how the Forest Service approaches evaluation. We gave our own summary and asked Claude to validate it.

By the end, we had a solid picture of how the Forest Service buys professional services. But every insight came with a lesson about what to look for and what not to confuse.

Turning the Process into a Skill

Here's where it gets fun. We asked Claude: We're thinking about using this kind of process on a regular basis. Can you create a skill?

If you don't know what a Claude skill is, it's basically text that describes a process, a set of instructions that Claude can follow in future conversations. Think of it as a playbook.

The skill Claude generated encodes everything we learned the hard way:

• Always request the tradeoff_process field explicitly — without it, you'll confuse solicitation procedures with evaluation methodology
• Pull competition details to get contract type (FFP, T&M, cost-plus)
• Start by resolving the agency to get the right org code
• Search across relevant NAICS codes — don't assume which ones the agency uses
• Check active solicitations for real evaluation language, because FPDS codes are approximations
• Pull protest history and actually read the digests
• Synthesize everything into a competitive intelligence briefing

Every dead end from the Forest Service conversation became a guardrail in the skill. Every wrong field request became a "make sure you use this field instead."

The Payoff: One Prompt, Any Agency

With the skill built, we pointed it at the VA: Tell me about agency procurement analysis at the VA.

It just worked. Claude read the skill, knew exactly what to do, and ran the full investigation without us having to guide it. It pulled the contract portfolio, checked trade-off versus LPTA across NAICS codes, pulled protest data, and synthesized the findings.

The VA analysis surfaced that SDVOSB is a significant competitive layer, that the agency leads with technical quality over price, that you need to know specific vehicles like T4NG, and that firm fixed price is the default contract type. All automatically. All from one prompt.

But while Claude is fine for research, you probably will still need to share the research with a stakeholder and inspect the data yourself. So we asked Claude to dump the sample data into a table with links. And it generated an interactive artifact right there in the chat: sortable, filterable, with direct links to USASpending for every contract. You could filter by NAICS code, see which contracts were LPTA versus trade-off, sort by contract type. And every record links back to the authoritative source so you can verify it yourself.

Doing It Live: NIH

Just to demonstrate we weren't making any of this up, we did one more query live, on camera. NIH. No preparation, no pre-loaded data.

Claude read the skill, resolved NIH in the federal hierarchy, pulled the contract portfolio, and started cooking. It pulled contracts with trade-off process and competition data, checked protests, searched active opportunities — all the phases the skill prescribes.

The results: $2.87 billion across a hundred contracts in a six-year period. Trade-off way bigger than LPTA. A fair amount of both firm fixed price and time and materials. And a bunch of sustained protests in 2020 — which turned out to be the CIO-SP4 protest wave. If you know NIH procurement, that immediately makes sense. The skill surfaced a pattern, and domain knowledge explains it.

It even generated a visualization showing the evaluation methodology breakdown and contract type distribution. All from one prompt, while we narrated on the recording.

This is running on Sonnet, by the way. Not even the most capable Claude model. The skill is doing the heavy lifting.

Why This Works

Two things make this possible, and they're both necessary.

Structured data. Tango gives Claude access to actual federal procurement records — not web search results, not press releases. Contract data with real PIIDs, real dollar amounts, real competition codes, pulled from FPDS and normalized into clean, queryable endpoints. You can't analyze evaluation methodology if you can't get the tradeoff_process field. Web search doesn't have this.

Encoded expertise. The skill isn't just "search for contracts." It's "search for contracts with these specific fields, because without tradeoff_process you'll confuse solicitation procedures with evaluation methodology, and here's what each code means, and here's how to interpret the 'Other' bucket." It's domain knowledge captured in a form an AI agent can follow.

Structured data without expertise produces raw results nobody can interpret. Expertise without structured data produces speculation. Together, they turn one-off research into a repeatable capability.

What This Means for You

The way people interact with procurement data is changing. The next analyst researching a recompete might not open a browser, they'll ask an AI agent. But that agent needs a structured, reliable data layer underneath it. That's what Tango provides, and the skill is what turns access into expertise. You encode the hard thinking once, and from then on, every agency analysis starts from a foundation of accurate, structured, authoritative data, not AI-generated summaries you have to take on faith.

If you're building a platform or tool in this space, this is the pattern we want to enable. You shouldn't have to build your own government data pipeline to power features like this. Tango is the infrastructure layer — the plumbing underneath your product. Companies with established platforms are already bringing Tango into their stack instead of maintaining their own ingestion pipelines. Whether you're building an AI agent, a capture management tool, or an analytics dashboard, the data layer is the same. We handle it so you can focus on what makes your product valuable.

If you want to try this yourself, here's how:

1. Sign up for a Tango account — there's a free tier
2. Connect Tango in Claude's settings under Connectors (just search for "Tango")
3. Or connect via MCP at govcon.dev/mcp with your API key
4. Ask it about an agency. See what you find.

The skill we built is available too. If you email us at hello@makegov.com, we'll be happy to share it. Then you can point it at any agency and get a full procurement analysis: evaluation methodology, contract types, protest risk, vehicle preferences, and what it all means for competing.

We'd love to see what you build on top of Tango. If you have questions, reach out at hello@makegov.com.