Running Your Search on AI: The Setup That Makes Claude Useful

Searchers do not have an AI problem. They have a memory problem.

Most buyers try AI like this: open a new chat, paste in a CIM, ask "what do you think?" The answer comes back generic, so they conclude AI is fine for emails but not for serious search work.

The problem is not usually the model. The problem is that the model has no memory of the search. It has not seen the last 12 deals you passed on, the add-backs you no longer trust, the brokers who send real opportunities, the industries you are warming up to, or the customer concentration you will not tolerate. Every chat starts from zero, so every answer is generic.

The goal is not to make AI smarter. The goal is to stop making it start over.

Here is the difference in one example.

Weak: "Here is a CIM. What do you think?"

Strong: "Score this CIM against my buy box, compare it to the last 12 deals I reviewed, flag anything that violates my hard-pass criteria, identify the top 5 diligence questions, and update the pipeline file with the next action."

The second prompt is not more clever. It is just standing on everything the AI already knows about your search, because that context lives in files it can read. This piece is how you build that: a free file setup any searcher can run by Friday, the three workflows worth automating once it works, and where a tool like Searcher OS lets you rent the parts you should not build. It covers the operations side of the search, how you run your own pipeline. The sourcing side, how you find off-market deals with AI, is its own playbook.

None of this needs special software. A capable model plus disciplined file hygiene gets a searcher most of the way.

Start with the desktop app. Install Claude Desktop and start with Cowork. Cowork lets Claude read from and write to local files you choose to share, which is the feature this entire workflow depends on. Cowork runs inside the Claude Desktop app and is available on paid Claude plans. Your search lives in documents, and an AI working directly against those documents beats copy-pasting into a chat box every time.

Build the folder. One folder per search. Every conversation about the search runs inside it. Here is the structure to copy:

What each part holds:

CLAUDE.md = who you are, what you buy, how you judge deals. CONTEXT.md = where the search stands today. /cims/ = every CIM, named consistently. /targets/ = one note per company. /brokers/ = one note per broker relationship. /outreach/ = email templates and live sequences. /financials/ = Excel files, tax returns, add-back notes. /lois/ = submitted and draft LOIs. /diligence/ = QoE, legal, insurance, lender, and customer notes.

CLAUDE.md is the file that compounds. Claude reads it first, every session. It is who you are. Here is a starter you can paste in and fill out:

If you want help sharpening the buy box, the 10 criteria for buying any small business is a good gut check before you fill it in.

CONTEXT.md is the file that carries state. If CLAUDE.md is who you are, CONTEXT.md is where you are. Update it at the end of every session so the next one picks up where you left off. Starter:

Three prompts to run on day one, once the folder exists:

"Read CLAUDE.md. Ask me 10 questions that would sharpen my buy box, then rewrite the buy-box section tighter."

"Here are 3 CIMs in cims/. Score each against my buy box and tell me which one is worth a real look and why."

"Update CONTEXT.md with what we did this session and the next 3 actions."

A clever prompt is a rounding error next to a model that has read six weeks of your search and knows your criteria cold. The setup above is what makes that happen.

Do not blindly dump every CIM, broker email, tax return, or seller file into an AI tool. Check the NDA on each deal. Use a paid or work account where possible. Limit the folder Claude can access to the search itself. Keep human approval on anything that leaves your outbox. AI should help you screen and organize deals. It should not become an uncontrolled forwarding agent.

The same caution applies to connectors. Depending on how a connector is configured, it can have read or write permissions, so treat connecting one like giving an assistant access to your inbox, CRM, or deal room. Anthropic's connector documentation notes that connected services may access or modify data based on your account permissions. Read what you are granting before you grant it.

And keep the boundary clear: AI is not your lawyer, your lender, your QoE provider, or your investment committee. It should not decide whether to buy a business. It should help you make fewer lazy passes, fewer emotional yeses, and fewer avoidable misses. The judgment stays yours.

Once the folder system works, three workflows are worth standing up. They run in Cowork. Searchers who want to automate more deeply can step up to Claude Code, the more technical agentic version, which reads a whole project, edits files, and runs commands from the terminal and inside code editors. Either way, each workflow below takes an afternoon and pays back for the rest of the search. Go deep on the ones that fit; do not try to run all three in week one.

Inbox drag: NDA requests, CIM follow-ups, broker replies. The early months of a search are mostly getting on broker lists and chasing documents. Connect Claude to your email and it can draft the NDAs and CIM requests that go to brokers, in your voice, ready for you to review and send. A broker replies, the next message is drafted and waiting, and you approve what leaves your outbox. The administrative drag of the first 90 days drops sharply.

CIM triage: a standardized first pass before human review. This is the highest-leverage build for most searchers, and the real version of AI for due diligence. Define a reusable instruction once, reading straight from your CLAUDE.md buy box, so every CIM comes back overnight in the same shape:

Business summary Revenue, EBITDA, and SDE Asking price and multiple Customer concentration Owner dependency Add-back quality, with the soft add-backs called out Working capital issues Financing feasibility Top 5 diligence questions A verdict: pass, maybe, or real look

Pair that report with a scoring rubric so the verdict is consistent across deals and across weeks:

The report is the gate. It decides which deals earn your actual review time and which get a fast no. It does not replace a quality of earnings report, which comes later and goes deeper. It is the screen that tells you which deals are worth paying for one. When reviewing hundreds of deals, a consistent first pass is not a convenience. It is the difference between a search that scales and one that drowns.

Pipeline memory: one markdown file per deal, always current. Keep deal state in plain files inside the search folder, one page per deal: stage, financials, open questions, last contact. Because it is the same files Claude is already working in, you can ask real questions and get real answers. Which deals have gone quiet. Which still fit after you tightened the buy box. What you are waiting on, and from whom. You get a working pipeline view without standing up and maintaining a separate CRM.

The folder system is worth building because it captures the part of the search that is actually yours: your criteria, your notes, your pattern recognition, and your record of why you passed on deals. That is the memory layer. It should compound over time, and it should stay close to the buyer.

The layer underneath it is different. Sourcing data, CIM intake, pipeline tracking, broker follow-up, SBA math, and site monitoring are mostly operational plumbing. They matter, but maintaining them by hand can quietly become a second job.

Josh Thacker, founder of Searcher OS, has seen this problem from both sides. He spent a year searching before building software for searchers, and his view is clear:

"Most searchers do not lose because they cannot find deals. They lose because they cannot process enough deals consistently. The best buyers build a repeatable screen, then protect their time for the few deals that deserve real work."

That is the right framing. The goal is not to automate judgment. The goal is to stop wasting judgment on work that should have been standardized: finding the deal, pulling the basics, checking the obvious risks, updating the pipeline, and teeing up the next action.

Keep your search brain in your folder. Rent or automate the plumbing where it saves time. The question is not DIY versus buy. It is which layer deserves your hours. For most searchers, the answer is the judgment layer, not the infrastructure layer.

Before you do anything clever with AI, do the boring thing first. Here is the whole on-ramp:

That is it. Most searchers never get past the blank-chat stage and conclude AI does not help, when the fix was never a better prompt. It was a folder.

The searchers who win will not be the ones with the cleverest prompts. They will be the ones whose judgment compounds every week because their search never resets.

Disclaimer: This guide is for educational purposes only and does not constitute legal, financial, tax, or investment advice. Business acquisitions involve significant risks, and outcomes can vary widely based on individual circumstances. Always consult with qualified professionals including attorneys, CPAs, and financial advisors before making acquisition decisions. The EBIT Community does not guarantee the accuracy of information provided or the success of any acquisition strategy. Past performance and examples do not guarantee future results.