Numbers you can
defend in the room.

Every forecast is built on top of one year's numbers. Get that year wrong and every projection after it inherits the error. Kestrel's reported profit of $3.1M contained three things that will not happen again, and one that was never really a cost of the business.

• The settlement: a one-time payment from Amazon, booked as if it were a sale. Real cash, but no customer bought anything.
• The capitalised content: $1.2M of photography and video treated as an asset on the balance sheet rather than a cost. It is marketing, and marketing is spent, not owned.
• The cash test: for every dollar of profit Kestrel reported, only 31 cents actually arrived as cash. That gap is the question. Here the answer is that the profit went into a warehouse full of stock.

Normalised — meaning stripped back to what repeats — the real figure is $1.4M, not $3.1M. Forecasting from the reported number would have made everything downstream wrong by more than double.

Kestrel spends an average of $71 to win one customer. That is its acquisition cost, and the average looks healthy. Split by channel it is $23 through organic search and $124 through paid social — on the same $184 average order.

• What a sale actually leaves behind: contribution is the money left after the costs that exist only because that sale happened — the product, the shipping, the payment fee. At Kestrel, $87 on a $184 order, or 47%.
• Returns change the answer: 28% of apparel comes back, and that cost lands on the channel that sold it. Paid social sells apparel, which is why a channel that looks fine on revenue turns out to lose money on contribution.
• Worth versus cost: at a ratio of 2.1, a customer returns roughly twice what it cost to win them. But that depends entirely on how long customers stay, and the prompt flags the retention assumption rather than letting the ratio stand alone.

The verdict is not “the unit economics work.” It is “they work on three channels out of five, and most of the budget is going to the other two.”

“We missed budget by 8%” is not an explanation. A variance bridge breaks that gap into its causes and puts a number on each, so the conversation moves from blame to arithmetic.

• The four questions: did we sell fewer units (volume), at lower prices (price), of the wrong things (mix), or did our costs move (rate)? At Kestrel: volume −$3.1M, price +$0.9M, mix −$0.2M — and the three sum exactly to the $2.4M miss, which is the check most people skip.
• The margin side: of the 13 points of margin lost, 5.2 came from discounting, 4.1 from freight and a supplier switch, 1.9 from mix, and 1.8 from simply selling less across the same fixed costs.
• One-off or structural: a one-off happened once and will not repeat. A structural cause is still there next quarter. The prompt calls the discounting structural — because the stock that forced it is still on the shelf.

It then tests the explanation the user arrived with. Kestrel believed it was a freight problem. Partly right, mostly wrong.

Kestrel sells 64 different products — SKUs, in retail language, each one a separately stocked item. Reported together they look like one business. Ranked by the profit each actually contributes, eleven of them earn 79% of it and fifteen lose money.

• Why fixed costs decide the swing: rent, salaries and the warehouse cost the same whether sales rise or fall. That is operating leverage, and at Kestrel it runs at 3.4 times — a 10% fall in revenue takes 34% off profit. Most businesses model that on the way up and never on the way down.
• The allocation problem: shared costs have to be charged to products somehow, and the method decides the answer. Charging the warehouse per unit flatters big heavy cheap items; charging it by the space each one occupies flips two more products into losses.
• Why that matters: profitability by product is always partly a judgement. The prompt states which basis it used and shows how the conclusion changes under the other, which is the honest version.

Verdict: a mix problem, not a pricing problem. Those need completely different fixes.

“We will grow 18% next year” is not an assumption, it is a result. A driver tree breaks that result into the handful of things an operator can actually influence, so every number in the forecast can be traced back to a decision someone makes.

• Kestrel's tree: revenue is visitors to the site, times the share of them who buy, times what they spend, times how often they come back. Change any one and the revenue changes; there is nowhere else for growth to come from.
• Which one decides it: the share who buy — the conversion rate. Moving it from 2.1% to 2.4% is worth $6.2M over the horizon. Moving the average order by the same proportion is worth a fraction of that.
• The credibility floor: below 1.9% conversion the whole forecast stops holding. Stated as a number, so it can be checked, rather than as a worry.

Three things then get watched monthly — acquisition cost, four-week repeat rate, returns — so the forecast breaks early and cheaply rather than late and publicly.

A company's finances are described by three connected reports. The profit and loss shows what was earned, the balance sheet shows what is owned and owed, and the cash flow shows what actually moved. Because they describe the same events, they must agree — and when they do not, an assumption is wrong.

• What broke: the model's closing cash came out $2.1M below the balance sheet. The prompt traced it to an assumption that stock would sit in the warehouse for 71 days, when the history implied 118.
• What most models do: quietly adjust a number until the two sides agree. The error then survives all the way into the board pack, where nobody can see it.
• What this prompt does: reports the break with the figure and stops. The disagreement was the useful finding, not an inconvenience.

It also predicted the challenge: why would stock move 40 days faster next year when it moved 47 days slower last year? Nobody had an answer.

Kestrel is profitable and nine weeks from running out of money. Those are not contradictory, and the thirteen-week view is where the difference shows up — because payroll happens on a date, not in an average month.

• The cycle that explains it: Kestrel pays for stock 118 days before it sells, collects from customers in 4, and pays its own suppliers in 38. Net, it funds 134 days of trading out of its own pocket. Every single day of stock is $131K of cash sitting still.
• The covenant: the bank's lending agreement requires a minimum balance — here $500K. Going below it is a breach, which hands the lender rights it did not have the day before. Kestrel bottoms out at $310K in week nine.
• One answer, not five: the prompt works through profitability, growth, working capital, debt and timing in order, and commits. Kestrel is not unprofitable and not growing too fast. It is financing a warehouse.

The breach is stated in the first line of the answer, not buried in week nine of a table.

$4.1M spent today buys money that arrives over years, and money later is worth less than money now. Discounting future cash back to today's value gives the net present value — here $1.9M, meaning the project is worth $1.9M more than doing nothing.

• The bar it has to clear: the hurdle rate, 12% in Kestrel's case, is the return the company could get elsewhere. Anything below it destroys value even if it technically makes money.
• Payback is the other half: 3.4 years to get the $4.1M back. A committee cares about both — a high return that takes eight years may be unaffordable regardless.
• The benefit that got excluded: “improved customer experience” had no owner and no measurement, so the prompt left it out of the calculation and showed the figure both ways, $1.9M and $2.6M. A benefit nobody is accountable for delivering usually is not one.

It also names the break point: this works as long as the throughput gain exceeds 71% of projection. More useful to a committee than a single number.

Most downside scenarios are comfortable ones. Sales fall, costs helpfully fall too, and the business survives on paper. In a real downturn the damaging things arrive together, and the prompt is instructed to move them together.

• Rank before you model: the drivers are sorted by how much each one swings the outcome, so effort goes where it matters. For Kestrel: conversion rate first, freight second, returns third.
• Break points, as numbers: not “margin pressure is a risk” but “below 38.4% gross margin we breach the lending agreement.” A number can be monitored; a worry cannot.
• Decide now, not later: if the four-week repeat rate is under 18% by March, the autumn order gets cut by 40%. Committing to the response in advance beats improvising it in the week it happens.

The prompt assigns probabilities to each scenario and then says plainly how little confidence to place in its own probabilities, which is the honest position.

Boards do not read financial statements, they read the story and then attack it. The prompt builds both halves, and the second half is the one people skip.

• One governing thought: the single sentence that survives if they remember nothing else. Here: Kestrel is a working-capital problem wearing a margin problem's clothes.
• Three numbers, chosen deliberately: eleven weeks of cash, 41% gross margin, 118 days of stock. Not a dashboard — three figures each with a comparison and a direction.
• Seven questions, ranked: by how much damage each does if it lands unanswered, including the ones that cannot be fully answered yet and how to handle those without spin.

It also names the single figure most likely to get challenged and prepares its full derivation. For Kestrel it guessed right — why did you buy the inventory?