Build from first ask to owned workflow

Never really used AI? Start here. You’ll talk to a real one, step by step — on the kind of work you already do — and walk away with the whole way of thinking. No jargon, nothing to break.

Now on a real deal: start hand-in-hand, finish alone — build complete asks (task, facts, shape) until you pull a forwardable recommendation on a deal you’ve never seen, solo.

Same facts, different words, a different answer. Four levers: the verb, the guardrail, the reader, the example. Then rescue a weak prompt and take the trick home in your own voice.

The answer’s close but not right. Feel the wall of asking again, then steer: add one line, zoom one angle, make it audit itself — and finish the gauntlet without ever restating the deal.

You told it to keep replies short. Then you opened a new chat and it forgot, so you typed the rule again, and again. Pin it once instead and watch it ride every message you send from then on, in every chat.

A cap rate cannot be worked out from a price and a gross rent, and you will still get one back. Feel that once, then make the check a standing habit: it names what it assumed, before you send the number on.

A fresh chat has no idea who you are, so the answer fits any broker in the country. Pin who you are and what you work on, and the same question comes back about your market, your deals, your desk.

The ask you built took five minutes and you rebuilt it from a blank box next month. Keep the skeleton and swap only the numbers, so a monthly job takes one edit instead of one rewrite.

Ask about a lease clause from memory and you get an answer about leases in general — useless for yours. Feel that once — then hand over the actual clause and get an answer from your own document, one you can point to and check.

A deal is never one document. Ground your question in only the lease and the answer can be right about that page and wrong about the deal — because an amendment changed it. Hand over the whole stack and make it reconcile which one governs.

The answer is good, and you still can’t send a word of it to anyone. Feel that once — then say who’s receiving it, and get back the finished thing: your own numbers out, the committee’s numbers in, nothing wrapped around it.

You just learned to hand over the whole document. This is the one about which parts were never yours to send — strip it to the terms, and watch the answer come back just as good with none of the exposure.

Stop reading fifty pages by hand. Build a tool across five milestones — hand over the real lease, name the standard abstract fields, make it flag the clause a fast read misses, pull the exact words before it goes out — then keep the whole ask and run it on the next lease that lands.

Stop running the rollover waterfall by hand. Build a tool across five milestones — hand over the real rent roll, name the standard views, make it flag the risks, check a number before it goes out — then keep the whole ask and run it on the next property that lands.

Stop cross-checking a broker’s glossy deck by hand. Build a tool across five milestones — hand over the real offering memorandum, name the in-place-vs-pro-forma split, make it flag the spin, check a number against the source — then keep the whole ask and run it on the next OM that lands.

Stop turning a pile of public links into a confident brokerage claim. Build a dated research desk instead: bound the Chicago Loop search, register each source and conflict, rank listing candidates and occupier signals separately by research readiness, hand the live check to a named research lead, then clear the chat and cold-refresh an LA/OC industrial brief without stale facts.

Stop turning a map pin into a conclusion. Build a complete pre-underwriting place packet instead: resolve the North Bethesda site identity, draw and prove a strict radius, validate nearby projects and map evidence at four scales, separate observations from inferences, hand the gaps to the Market Intelligence Director, then clear the chat and cold-run an industrial site without stale facts.

A parcel shaded green on a map is not an eligibility determination. Screen a LIHTC candidate against the plan year’s thresholds instead: register the policy version and every dataset date, run each rule exactly as written to PASS, FAIL, or HUMAN REVIEW REQUIRED, keep a measurement to what it proves, write the narrative last, then hand the unmeasured distance and the overlay conflict to the Affordable Housing Policy Director and cold-run a second county under a newer plan.

A headshot generator hands back twenty-four polished portraits in two minutes and answers none of the questions that decide whether any of them can be published. Build the governance layer instead: separate the vendor license from the signed consent, grant only the uses the form version actually named, measure every generated likeness against the person’s own reference photos, put a date and an owner on every stored copy, then hand the release to the Brand and Advertising Compliance Director and run it cold on a second firm whose vendor contract is inverted.

A deal you’re inclined to recommend, and an answer that agrees with you — the dangerous kind. Take a real acquisition to committee: make it argue the case to pass, route the legal question to counsel, sign a recommendation that carries your own call, then keep the pass-test kit and run it cold on a second deal. The part you can’t hand over.

You inherit a colleague’s underwriting model and it goes to committee Thursday under your name. Ask it to check the file and it hands back a clean bill of health — correctly. Build a model auditor across five milestones instead: put your firm’s standard and the source documents in front of the work, overrule the one finding that isn’t an error, write the note back to the person who built it — then write a sixth rule of your own into the standard and run the whole kit cold on a second inherited model. You keep the kit. Rule seven comes from the next model you check.

A familiar anchor, an open store and a renewal conversation turn into an unsupported “investment-grade” answer. Build the procedure the shortcut skipped: gate five kinds of evidence, show every ratio and lookup, separate a house grade from an agency-issued rating, record the named human decision, author your own stop rule, then clear context and run the kit cold on a stronger tenant with one critical gap still open.

A polished heat map turns mismatched sources and correlated risks into confident scores, stacked penalties and an unsupported exit-cap change. Build the judgment procedure underneath it instead: lock the decision, register source scope and conflicts, anchor likelihood/severity/trajectory, cluster shared drivers, authorize only evidenced model changes, add your own stop rule, then clear context and run the kit cold on an industrial lease-up.

Ask an AI tutor whether you got it right and it will tell you well done — which is how a wrong formula becomes a signed-off skill. Build the loop that cannot hand out progress it did not verify: a baseline task instead of a self-rating, one cited concept at a time, an evidence bar written before the ruling, a binary gate that stays shut on confidence, a retry aimed at the actual defect, and a dated ledger that leaves the certification to the Head of Underwriting Standards. Then run the same loop cold on a standard with no spreadsheet in it.

Stop reintroducing your desk every fresh workspace. In VS Code with Claude Code, write the project instructions once — who you are, the first-pass screen you want, the source-backed math check, and the legal boundary — then cold-run the same setup on the next deal with one line.

Stop screening one file at a time and stitching the answers together afterward. Point Claude Code at a folder of real industrial transaction packets, force every deal into one source-tied comparison board, catch the ownership-basis trap — then clear the session, add a fourth folder, and refresh the whole board with one line.

Your own verification rule passes a figure whose only source is a broker’s email — the rule fired exactly as written and the number is still going to committee. You wrote it against a folder that had no correspondence in it. Amend it to tell a document of record from somebody’s assertion about one, then open a repo whose project instructions somebody else wrote and work out, before you touch anything, what it forbids, what it assumes, and what it is silent about.

Stop exporting a fresh diligence room every time one document changes. Give Claude Code a project-scoped, read-only path to the live source, prove the access boundary, reconcile a base lease with its amendment, then clear the session and catch a new estoppel with one-line refresh.

Stop rebuilding the valuation in a spreadsheet every deal. Here the sandbox becomes VS Code with Claude Code inside it — drive it across five milestones to normalize the rent roll, ground the assumptions in the deal file, build the stabilized pro forma, land the value, then wrap the whole pipeline as a skill you re-run on the next property by name.

A sourcing agent hands you ten matched properties, ten broker contacts and a route link — and nothing on the screen tells you which building is listed twice, which listing went stale a year ago, which email was guessed rather than verified, or which tenant another brokerage already represents. Build the pipeline that keeps hard gates apart from preferences, dedupes on identity, splits verified contacts from inferred ones, orders a tour without depending on one vendor’s link, and leaves every outreach for the Managing Broker to approve. Then cold-run a new market with a delta report.

One deal card, three completely different kinds of fact: a public record with a date, a cell in your own model, and something a human typed. Build the desk that keeps them apart — because the Returns Bar is showing green off a tagged cell that moved when someone inserted a row, reading a value the workbook cached and never recalculated. Then port the whole thing to a county whose assessor is missing a field.

A one-page valuation comes back looking finished — a single number, a confident paragraph, six comparable sales. One is a sale-leaseback at above-market rent, one closed twenty-six months ago, one is an asking price, one has a quarter of the term remaining, and the tenant is called investment grade rated when nobody rated it. Build the pricing book instead: eligibility before adjustment, every adjustment and exclusion on the record, a range with its drivers, and a conclusion that stays pending until the Head of Net Lease signs it.

The dashboard says market rent is up four percent and the team is about to raise renewal asks on it. It rose because the sample shifted to bigger, newer units, and because asking rent stopped meaning anything the month concessions arrived — net effective rent actually fell. Build the signal that carries its own methodology receipt: a stated universe, a normalized unit, asking held apart from effective, measured coverage and bias, a real seasonal comparison, and a backtest against the leases that actually signed. Then port it to a market with no university cycle.

The budget dashboard says eight percent contingency remaining, on budget. It is already gone: a change order approved by email but never executed, a dollar counted twice because the commitment and its invoice both landed in the same draw, and retainage recorded as spent when it is still withheld. Build the control tower that reconciles budget, commitment, spend and exposure to one number — and blocks a payment to a vendor whose insurance lapsed mid-draw. Then run next month’s draw with a credit memo in it.

The generator hands back eighty-four items and it reads exhaustive. It is also structurally wrong — it assumed a fee-simple purchase on a deal that is ground-leased, so the estoppel, the leasehold endorsement and the lender consent are simply absent, and length read as diligence. Build the tracker instead: every item with an owner, a status, the evidence that satisfies it and a date; conditions precedent separated out; a servicer timeline that makes the target close date impossible; and legal sufficiency left to Transaction Counsel. Then restructure it when the financing changes mid-stream.

The exported site-visit PDF looks complete — thirty-four items, twenty-one photos, a tidy summary. It is not reconstructible: two photos are one capture a sync retry duplicated, four carry the time they uploaded rather than the time they were taken, one silently inherited the previous photo’s location, and a note was edited at the desk three days later with no trace. Build the chain of custody instead, from a phone with no signal to a handoff nobody can quietly change — and get the blocked fire exit escalated the same day rather than filed.

Stop re-typing the same rent-roll ask into a fresh chat every deal — and watch it come up thinner every time. Drive Claude Code to write the workflow down once as a skill: name it, spell out the steps, then run it by name on the next deal in a single line. Your first tool, and you own it.

A historic deal library says it imported 377 records and still cannot find one known transaction. Build the source-tied database correctly: discover all 378 files, reconcile 375 canonical reports plus three stubs, write the ingest as a reusable skill, query a five-deal industrial cycle screen, then clear context and version a cross-era update without rewriting history.

A polished amortization app in twenty minutes, and its month-one interest is wrong — the reimplementation used actual over 365 where the workbook uses 30 over 360, and the chart looks superb either way. Build the discipline that makes a reimplementation trustworthy: freeze the verified logic and its test vectors first, keep the engine and the interface apart, and refuse to ship anything that does not match the workbook to the cent. Then migrate a model version cold, regenerating the vectors instead of nudging the engine until the chart looks right.

The generated case reads beautifully and cannot be solved: the narrative says ninety-two percent occupancy, the rent roll says eighty-eight, there is no exit cap anywhere, and the answer key uses a management fee that appears nowhere in the packet. Build the factory instead — a specification that declares its intentional gaps up front, a solver that works only from the published packet, and an adversarial checker that recomputes every stated total from its parts. The generator and the solver never share what they know.

Three good lease-review skills still make a stale memo when their handoffs live in someone’s head. Write one source-tied run contract, compose the skills in order, make every output stop at human approval, then register a durable Routine and prove a new commencement certificate refreshes the work from a cold session.

An analyst crafted a clear recommendation, but the model-drafted committee memo hides the fact that Bull and Bear disagreed fundamentally. Build the independent-voice procedure instead: four personas with opposed biases (bull, bear, skeptic, operator) each deliver independent verdicts, no averaging. Catch when synthesis smooths disagreement into false consensus. Author your own chairman's decision rule. Prove the procedure travels to a different deal class cold — the committee stays fixed, the risks and your verdict change.

Turn a green loan snapshot into a durable tenant → lease → NOI → loan surveillance system: bind each fact to its source and clock, carry one driver through the chain without double counting, preserve the immutable November run, then clear context and append a cold December refresh without letting the workflow make the human call.

The committee is about to read a memo describing a deal nobody is underwriting: it says a 6.25 percent exit cap, the model has said 6.00 since someone changed it two days later, and nothing flagged the drift. Build the governance instead — roles that decide who may change what, a change log for the model and a separate one for the narrative, review comments that cannot be closed without an answer, and a release package whose cited version matches what it actually attaches. Then handle a revised model with a delta release rather than a silent re-issue.

Most permit rejections are paperwork, not design — and the confident response package in front of you earns another one: a citation to a superseded code edition, an answer written against a sheet that is not in the submitted set, and a sentence asserting a structural detail is adequate, which is a design representation nobody here is licensed to make. Build the room that traces every claim to a dated authority and every response to its exact sheet, routes a state-versus-local conflict instead of settling it, and stops at the architect of record.

The final transfer test: no lease pipeline and no step-by-step recipe. Take an unfamiliar property AP batch, discover its contract/budget/ledger controls, build the reusable procedure, own the release queue, schedule it, then cold-handle a credit memo without turning automation into payment authority.