Build your own deal-review analyst.

What you'll end up with

Not a chatbot that has opinions about your deals — a small team with jobs. By the end of this walkthrough you'll have:

An analyst is only as good as what it has read. Yours gets two libraries.

a) Public: the investors whose judgment you trust

Pick a thinker whose framework you'd actually want applied to your deals — a well-known SaaS investor's public essays, a value investor's shareholder letters, a real-estate operator's talks. Have Claude gather what they've published — essays, conference talks, podcast transcripts — and then distill it: strip the banter and the politics, keep only the investing content, and compress it into paraphrased "framework cards" the agent reasons from.

Two rules make this work:

• →Set a source-precedence hierarchy. The primary thinker's own writing is the spine; supporting voices defer to it. Without this, the agent blends five people into a mushy average — or worse, fabricates a position nobody holds.
• →Spend where it matters. Bulk transcript-crunching can run on cheap or local models; save the strong model for judgment and synthesis. Hours of audio costs almost nothing to distill this way.

One line on IP: your private corpus is for your use. If you ever share your setup with a friend, ship the method and your scripts — not other people's verbatim content.

b) Private: your own deal history

Past decks, your old deal memos, term sheets, the deals you did and — just as valuable — the ones you passed on and why. Drop them in a folder. This is the part of the corpus that makes the analyst yours instead of a generic one.

# a corpus is just folders — no database, no magic
deal-review/
  corpus/
    essays/            # the thinker's own writing (the spine)
    talks/             # transcripts, distilled to framework cards
    my-deals/          # your decks, memos, term sheets, passes
  kb/
    rubric.md          # how every deal gets scored
    filters.md         # your standing preferences (grows over time)
    calibration-log.md # dated record of what you taught it

Here's the honest problem: most investors don't have a written pass/reject history. Your taste lives in your head. So instead of trying to write it all down up front — you won't — the agent builds it forward, one reviewed deal at a time.

After every review, the agent asks you two to four one-tap questions:

Thirty seconds of your time per deal. The leverage is in the promotion ladder: a one-off answer stays in the calibration log; a pattern that recurs two or three times gets promoted into an active filter the agent applies to every future deal; and when you teach it a threshold — the point where a metric goes from acceptable to alarming for you — the rubric itself gets updated, with a dated note. Review #30 knows your taste in a way review #1 couldn't.

The center of the team is the deal reviewer: a thin agent persona pointing at the rubric and corpus you just built. The design rule that makes it dependable is a fixed output contract — every review, no matter the deal, comes back in the same shape:

Around the reviewer, a diligence team fans out in parallel when a deal clears the first read: one agent sizes the market from primary sources, one runs the competitive scan, one backgrounds the founders or the sponsor, one stress-tests the financial model's assumptions, one compares the term sheet's terms against what's market. Each comes back with a sourced brief; the reviewer synthesizes. What used to be an afternoon of tabs is twenty minutes of reading their reports.

If your deals are buildings rather than companies: the same pattern powers property work — lease abstraction across a rent roll, acquisition diligence checklists, LOI drafting. We wrote a full walkthrough for that world: Claude Code for commercial real estate.

This is the step most people skip, and it's where the compounding lives. Don't let the agent invent its own report format — hand it the actual spreadsheets you already use: your diligence checklist workbook, your deal-comparison matrix, your portfolio tracker. Claude learns each one's structure and fills it in per deal, for your review — same format every time, so deals stay comparable across months and memory.

Four rules keep this honest:

Start from our kit

You don't have to build the skeleton from scratch. We open-sourced the structure we use ourselves — the agents, the rubric and interview-protocol templates, the safety rails — with zero data inside. Bring your own credentials, add your own corpus.

# inside Claude Code — two commands, then say hello to your CFO
/plugin marketplace add uwskiguy/claude-finance-team
/plugin install