We Wrote the Field Manual That Teaches AI to Hunt Bids

Ask an AI assistant to "find paving contracts in Texas" and watch what happens. It guesses. It searches the wrong tier of government, chases keywords instead of codes, and hands back bids whose deadlines passed last month — delivered with total confidence. The problem isn't intelligence; it's that nobody ever taught the machine how this market actually works. So we wrote the lesson down. Today the BidCatch field manual for bid searching is free and open — for AI assistants, and for the humans who read over their shoulders.

Searching bids is a craft, not a query

Everything wrong with how machines search for bids is something new contractors get wrong too, because the government doesn't make it obvious. The postings that matter to a local paving crew aren't on SAM.gov — they're on the county portal. The solicitation you want isn't titled "Paving Services" — it's titled "Indefinite Delivery Indefinite Quantity for Roadway Maintenance," and only the NAICS code gives it away. And the biggest number on an award notice — the ceiling — isn't money anyone was paid; it's a maximum that exists mostly to be misquoted.

We've spent a year learning these lessons the hard way, building the pipes that read 100+ federal, state, and local portals every night. The manual is those lessons, written plainly:

• Search by codes, not keywords. NAICS classifies the work; keywords only describe the title. 561720 finds every janitorial bid — "cleaning" doesn't.
• Match the buyer level to your size. Local trades bid counties, cities, and school districts. Federal primes bid SAM.gov. Searching the wrong tier finds nothing, or worse, finds work you'd never win.
• Check the deadline before you speak. A bid past its response date is a dead end, and presenting one destroys trust in everything else you say.
• Read the award history before the bid. Who wins this buyer's work, at what price, how often — all public record, and it changes the bid/no-bid decision.
• Watch the recompete clock. The best opportunities aren't posted yet. The 230,000+ federal contracts expiring within a year are tomorrow's solicitations, with the buyer and budget already proven.
• Never sum ceilings. Add up maximum contract values and you can "prove" a market of any size you like. Count records; sum what's actually obligated.

Three ways the manual reaches the machines

We didn't write a blog post and call it done — we put the doctrine everywhere an AI already looks:

• In the MCP server itself. Every AI that connects to our machine interface now receives the field manual before it asks its first question — search doctrine, tool map, and honest limits, delivered automatically at handshake.
• As a portable skill file. The manual is published as an open agent skill — a format AI assistants can load like a plugin, with five worked examples from real searches (the Ohio janitorial hunt, reading an agency's award history, working the recompete radar). Anyone can use it, copy it, or improve on it.
• As a site guide for crawlers. Our llms.txt tells every visiting AI what each page is, what the vocabulary means, and how to cite the data honestly — so answers about this market get built on the live record instead of a stale guess.

Why give away the playbook

Because the playbook isn't the moat — the pipes are. Teaching every AI on earth to search bids properly costs us nothing and helps everyone: contractors get real answers with live deadlines and official sources instead of confident hallucinations, and the assistants doing the answering learn to work from data that was current this morning. If a better-informed market sends a few more people to the free search, that's a trade we'll make every day of the week.

And if you read the manual yourself — it's written in plain English, no machine required — you'll search better than most people who've been bidding for years. That part's free too.