Prompt Pilotby Write.law
All prompts

Pressure-Test

Pressure-test your damages number

Redoes the arithmetic, ranks the assumptions by how far they move the total, runs the other side's version of each input, and hunts the double counts and the categories that are computed perfectly and not recoverable.

About 25 minadvancedLitigation, Insurance

Your prompt5,986 characters

Still to fill in: The damages model, The claims and which theory each category runs under, Governing law and forum

RoleYou are the lawyer who has watched a damages case collapse on a cross-examination about a single input, and who therefore adds the numbers up personally rather than trusting the spreadsheet somebody else built. You separate whether a figure is computed correctly from whether the category is recoverable at all, because a perfect calculation of something unavailable is zero. You will say when the number a client wants to put on a demand is not the number you would put in front of a factfinder.What I needAttack the damages model below on the claims as pleaded, under Governing law and forum, and tell me what the number survives as.InputsThe model: The damages model Claims and which theory each category runs under: The claims and which theory each category runs under Governing law and forum: Governing law and forum What supports each figure: What supports each figure Where this number is going: a demand letterHow to work this1. Rebuild the arithmetic from the inputs and add it up yourself. Say plainly whether the stated total is the total the inputs produce. For any figure that cannot be derived from what I gave you, say so and do not accept it; a number whose provenance nobody can reconstruct is the one that gets asked about first. 2. List every assumption behind every input and rank them by leverage: which one, if it moves, moves the total most. Report the ranking as a sensitivity, so I can see where the case actually is rather than where the work went. 3. Run the other side's version. Take each assumption to the most defensible adverse value, not the most extreme one, and show the total that results. Report a range with the two or three inputs that produced it, not a single alternative number. 4. Hunt the double counts and the gaps specifically: the same loss recovered under two categories, an item recovered under a theory not actually pleaded in The claims and which theory each category runs under, consequential damages where the contract excludes them, prejudgment interest computed from a date the claim did not accrue on, a mitigation offset that is missing, and a category that overlaps with an insurance recovery or another payment already received. 5. Separate recoverability from calculation, category by category. A category that is computed perfectly and is not recoverable in Governing law and forum is zero, and that is the most expensive kind of error in a damages model. Do not state the recoverability rule from your own knowledge: mark each one [CONFIRM - recoverability under governing law] and say what would settle it. 6. Say which figures rest on a document, which on a witness who will be cross-examined, and which on an expert who has not yet been retained or disclosed. A number resting on an undisclosed expert is not yet a number. 7. Give two figures and the reason they differ: the number I would put on a demand and the number I would put in front of a factfinder.Ask me firstYou have the model, the claims, the governing law, and the support. Ask me these three, which none of that reveals, then stop and wait: 1. Has anyone outside the case been walked through this number, and could they repeat it? A total that needs four slides to explain loses to a smaller one a person can hold in their head. 2. What is the other side's damages story likely to be, and have they disclosed anything: an expert, a counter-model, a document showing a different margin? 3. Is there a number the client has already said out loud, in a board deck, a press statement, an insurance claim, or an earlier demand? That number is now a ceiling or a floor and it will be found. Do not begin until I answer. If I tell you to proceed anyway, state each assumption at the top of your output and mark it [ASSUMPTION - verify].Output formatOpen with the arithmetic verdict in one line: does the model add up, and if not, what the inputs actually total. Then the sensitivity ranking: assumption, its current value, its leverage on the total. Then the adverse run with the resulting range and the inputs that drove it. Then the double-count and gap findings, quoted against the model. Then a recoverability table: Category | Computed correctly | Recoverable [CONFIRM] | Supported by. Then the demand number and the factfinder number with the reason for the gap. Close with one line naming the two of my answers that most changed this, and what number you would have landed on without them. If an answer changed nothing, say so, because it means I should not have been asked.Never do this- If the critique would apply to any damages model in any case, it is too generic. Every finding has to attach to a line in the model I gave you. - No hedging filler. "The lost profits figure is arguably aggressive" is not a finding. Say what the assumption is, what its adverse value is, and what the total becomes. Do not tell me to consult an attorney; I am the attorney. - Never state what Governing law and forum allows as recoverable, what interest rate applies, from what date interest accrues, or what a court has held. Mark each [CONFIRM] and cite nothing. Never invent a figure, a margin, a rate, or a date to complete a calculation. - Where a figure's basis is not in my inputs, say you do not know where it came from rather than reconstructing a plausible derivation for it. - Do not pad. If the model has one real vulnerability, say so and spend the output on it. Length is not value.Before you answer- Did I actually add the numbers, and do my totals reconcile? - Did I invent any figure, rate, or date not present in the inputs? - Is every category assessed twice, once for calculation and once for recoverability? - Did I take the adverse inputs to their most defensible values rather than their most extreme? - Would this critique be useless against a different model? It should be.

The run walks turn one, the pressure test, the follow-up, and a check on what came back. The Cockpit adds driver's-seat tunes. Your values come with either one.

2

Pressure-test it

Makes the AI switch hats and attack its own answer.

Their expert has your model, your inputs, and a deposition date. Take that chair as the expert being paid to take it apart: which single input would you spend the whole morning on, which figure has no document behind it that you could make the witness admit, and what total would you put in your own report using only the numbers my side produced? Then tell me what I fix before the disclosure rather than after it.
3

Go deeper

Pushes the work further once the basics are right.

A number nobody can explain out loud does not survive the room. Write the damages story: the sequence in which these figures get presented so each one follows from the last, the single sentence that carries the total, the one document that anchors it, and the answer to the question a factfinder always asks, which is why this number and not a rounder one.
4Check what came backPaste the answer here and work a checklist against this prompt's own rules.

Before you run it

What to gather first

  • The model itself, with its inputs and its arithmetic, not just the total
  • The claims actually pleaded and the theory each damages category runs under
  • What supports each figure: a document, testimony, an expert
  • The governing law and forum, because recoverability is where categories die
  • Whether this number is for a demand, a mediation, a disclosure, or a verdict form

Watch for

  • What is recoverable, how interest accrues, and from what date are forum-specific and often dispositive. The model will state all three confidently and wrongly. Confirm each.
  • The model can produce arithmetic that looks checked and is not. Reconcile every total yourself before any figure goes into a demand or a disclosure.
  • A number the client has already stated somewhere, in a board deck, an insurance claim, or a press release, becomes a ceiling or a floor. Find it before the other side does.
  • Double counting across theories is the most common defect and the hardest to see, because each category is defensible on its own.
  • Damages analysis is core work product. Do not paste privileged valuations or settlement positions unless your firm's AI policy and the engagement terms permit it.

What comes back

The arithmetic verdict in one line, a sensitivity ranking of assumptions by leverage, the adverse run with a resulting range and the inputs that drove it, double-count and gap findings quoted against the model, a recoverability table separating correct computation from availability with every rule marked [CONFIRM], the demand number and the factfinder number with the reason they differ, and the answer-consumption line.

See an example of what you’ll get
*(After I answered: nobody outside the case has heard the number; they have not disclosed an expert but their answer denies the margin figure; and the client claimed "approximately $1.5 million" in an insurance notice last year.)* Arithmetic verdict. The model does not add up. $1.42M + $842K + $61K = $2,323,000, not $2.39M. The $67,000 difference appears to be prejudgment interest folded into the total without being listed as a line. Break it out or remove it; a total that cannot be reconstructed from its own line items is the first thing a cross-examiner finds. Sensitivity, by leverage | Assumption | Current value | If it moves | Leverage | |---|---|---|---| | Per-unit margin | $79K/month | Each $10K of margin moves lost profits by $180K | Highest. The whole lost-profits category rests on one CFO declaration | | Lost-profit duration | 18 months | Each month removed is $79K | High. Nothing in the inputs explains why 18 and not 12 | | Cover was commercially reasonable | Assumed | Fails entirely if not, taking $842K to whatever a reasonable cover would have cost | High, and binary | | Storage costs are consequential | Assumed recoverable | Zero if the section 11.4 waiver reaches them | Binary and probably fatal to this category | | Interest accrual date | 4/14 | Moves the interest line only | Low in dollars, high in credibility | The adverse run. Taking each input to its most defensible adverse value rather than its worst case: margin at $58K (their likely position from the answer's denial), duration at 12 months, cover reasonable but at the lowest of the three invoices, storage excluded by the waiver, interest from the date of the complaint rather than the breach. Lost profits become $696K, cover becomes roughly $610K, storage becomes $0. The adverse total is roughly $1.31M against your $2.32M. The range is driven almost entirely by two inputs: the margin figure and the storage category. Double counts and gaps - Double count. Lost profits and cover costs both run 18 months. If the cover replaced the supply, the lost profits for the covered period are the margin differential, not the whole margin. As modeled, you are recovering the profit you would have made and the cost of making it. This is the finding that most threatens the number's credibility, because it is defensible line by line and indefensible when the two lines are read together. - Gap. No mitigation offset appears anywhere, and the model does not state that none was available. Silence reads as an omission. - Theory mismatch. The tortious interference claim is pleaded but no damages category runs under it. Either it supports a category you have not modeled or it should come out of the damages narrative entirely. - Interest. Computed from 4/14, the repudiation date. Whether the claim accrued then or on the date of each unfilled order is a question the model answers silently. [CONFIRM - accrual date for prejudgment interest under governing law.] Recoverability | Category | Computed correctly | Recoverable | Supported by | |---|---|---|---| | Lost profits $1.42M | Arithmetically yes, subject to the double count | [CONFIRM - governing law, and whether the contract's limitation reaches lost profits] | A CFO declaration. No document. No expert. This is a witness, not a number | | Cover $842K | Yes | [CONFIRM - and confirm the reasonableness standard] | Three invoices, Ex. C-1 to C-3. The strongest figure in the model | | Storage $61K | Yes | [CONFIRM - section 11.4 consequential waiver. If it reaches storage, this is zero] | Not stated in the inputs | | Interest $67K | Not shown as a line at all | [CONFIRM - rate and accrual date] | Nothing | The two numbers. On a demand, $2.32M, stated as line items so it can be reconciled, with the storage line separated so that conceding it later costs nothing. In front of a factfinder, the number I would actually argue is closer to $1.6M: cover at $842K plus lost profits limited to the margin differential for the covered period, with storage abandoned and interest pleaded separately. The gap between the two is not a negotiating tactic. It is the double count and the waiver, and if the demand number goes out without the line items, the trial number looks like a retreat instead of a discipline. The number the client already said. You told me the insurance notice said "approximately $1.5 million." That is now a document the other side will find, and it sits almost exactly on the number I would argue at trial. That is either the best fact in this analysis or the worst, depending on whether the notice explains what it covered. Get it and read it before any demand goes out. What your answers changed. The insurance notice is why there are two numbers in this output rather than one, and why the demand is itemised. Without it I would have recommended a single $2.32M demand and you would have been handed your own $1.5M figure at mediation. That they have denied the margin figure in the answer is why margin sits at the top of the sensitivity table and why the adverse run uses $58K rather than a guess; it also means the CFO declaration is not going to be enough and the expert question is now urgent rather than optional. Your answer that nobody outside the case has heard the number changed nothing analytically, though it is the reason the expand step exists, so it earned about half its round-trip.
Why this prompt is built the way it is
## Framework 1. **Redo the arithmetic.** Add it up. Say whether the total is the total, and name any figure that cannot be derived from the inputs. 2. **Rank the assumptions by leverage:** the one that moves the total most if it moves. 3. **Run the adverse version** of every input at its most defensible value and report the resulting range. 4. **Hunt double counts and gaps:** one loss under two theories, an item under an unpleaded theory, excluded consequentials, interest from the wrong date, a missing mitigation offset. 5. **Separate recoverability from calculation,** and never state a recoverability rule from memory. 6. **Say which figures rest on a document, which on a witness, and which on an undisclosed expert.** 7. **Give the demand number and the factfinder number, and why they differ.**