One Claude Prompt, Full Underwriting: What Has to Be True First

Technically, yes. A workflow that runs intake, normalization, assumptions, model population, and IC drafting in sequence from a single instruction is buildable today, and PSV builds them. The pattern that works is not one giant prompt but a chain of small, named steps, each with defined inputs and outputs, composed in sequence. This matches the engineering consensus Anthropic itself publishes for agent design: simple, composable workflow patterns are more reliable than a single autonomous agent improvising across a long task. Each step does one job, hands a checkable artifact to the next, and fails loudly instead of wandering. The single prompt is just the ignition. The engine is the chain, and every link in it existed before the button did.

What one prompt buys is the removal of human glue between steps a firm has already proven individually. That is worth real hours: the analyst stops re-uploading files, re-explaining context, and re-formatting intermediate outputs five times per deal. What it does not buy is any improvement in the steps themselves. A normalization step that misreads concession structures produces the same misread inside a chain, except now the error flows silently into the assumptions memo, the model, and the IC draft, wearing the formatting of finished work. Chaining multiplies both the leverage and the blast radius. That is the whole reason the prerequisites below matter more than the chain.

First, the files have to be findable and consistent. A chain cannot ask a person where the current rent roll lives, so the deal folder needs one canonical location per document type and a naming convention the workflow can rely on. Second, the arithmetic has to leave the language model. Every number that matters, in-place rent, loss to lease, expense ratios, debt service, is computed in code the firm can read, with the model doing extraction, labeling, and prose. A language model that computes a DSCR inline is a liability no reviewer can audit. Third, every assumption gets an owner. Exit cap, rent growth, downtime, insurance: the chain can propose defaults from the firm’s standards file, but a named person accepts or overrides each one, and the memo records who.

Fourth, the review gate has to survive the automation. The seductive failure mode of end-to-end chains is that output arrives looking done, and organizations quietly stop reading what looks done. The NIST AI Risk Management Framework’s core insight applies directly: managing AI risk is about governance and measurement around the system, not just the system’s accuracy. In practice that means the chain ends at a draft with a diff-friendly assumptions page, a named reviewer works through it against the source documents, and nothing reaches a committee or a counterparty until that sign-off exists. The chain saves the hours between steps. The review hour it must not save is the one at the end.

Run the chain per deal, but adopt it per stage. The rollout that works starts with normalization alone until the team trusts its output on ten real deals, then adds the assumptions stage, then the model, then the draft. Each stage earns its place in the chain by being boring, and the one-prompt version arrives last, as a convenience wrapped around five proven steps rather than a leap of faith. The inputs, outputs, and gates are in the table above. The part worth underlining is the failure behavior: a well-built chain stops and reports when a document is missing or a number does not tie, rather than improvising past it. If a vendor demo never shows you what the chain does when the rent roll is a scanned PDF with a handwritten column, ask.

Measure the thing you actually wanted. The point was never document generation speed, it was cycle time from deal receipt to a screened yes or no, and error rate at the IC stage. A firm that tracks those two numbers before and after adoption knows what the chain is worth and where it leaks. In PSV’s experience the chain’s value concentrates early, screening more deals to a fast, documented no, and the review gate’s value concentrates late, on the deals that go to committee. Both numbers belong to the firm, not the vendor, and both should be visible to the people signing the memos.