The Decision Frame
Ten chapters of instinct deserve one card.
Tessa's is a postcard pinned above her desk, four questions long, and half the office has copied it. It is not a policy and nobody signed off on it. It is the thing she found herself asking, in the same order, every time somebody arrived at her desk with a task and a hopeful expression — written down so she would stop having to remember it.
What follows is that card, one question at a time, with the chapter that taught each one named as it goes. This is the whole course folded to pocket size. Nothing on it is new.
Q1: What Does Being Wrong Cost?
First, always, and it is the stakes question from Chapter 3, promoted from a habit to a gate.
There are three honest lanes, and almost everything falls cleanly into one of them.
Wrong is cheap. Thinking out loud, brainstorming names, drafting notes only you will read, getting unstuck on a blank page. If the answer is wrong you find out immediately and nobody else is affected. Use the model freely here, and stop apologizing for it.
Wrong is embarrassing and costs money. Anything a guest, a client or a supplier will read: an itinerary, a quote, a policy summary, a brochure paragraph. Draft with the model, then verify the facts the thing stands on before it leaves your hands.
Wrong is serious. Legal, financial, medical, contractual, safety. The model may draft at most, a qualified human owns every fact in the result, and where that ownership cannot actually happen, the task stays out entirely.
Notice that Q1 never says no on its own. What it does is set the price of the rest of the card — because the higher the stakes, the more the next question has to deliver.
Q2: Can I Check It, Affordably?
Second, and this is the question that does the real work, because it is where good intentions go to be tested.
Verification has to fit. Not exist in principle — fit, into the actual workday, at the actual volume. This book has shown three shapes that fit, and you have used all three.
A fact you can confirm in two minutes: the load-bearing claim from Chapter 3, and the two minutes on a hotel's own website that was the entire Harbourview defence. A sample you can read: Chapter 5's random dozen rows opened against the reviews they came from, twenty minutes for a file of 2,300. A check that runs by itself: the rules in Tessa's program that test shape and range on every row, and the tests in Chapter 9 that do the same to software.
Now the part that makes the card honest instead of decorative. If no check fits — if the only way to know whether the output is right would cost more than doing the work yourself — then read Q1 again. In the cheap lane, fine: an unchecked brainstorm harms nobody. In the serious lane, "I could not check it" and "it went out anyway" cannot both be true. The frame closes that door rather than leaving it politely ajar, and that is the closest this card comes to having an opinion.
Q3: Is Judgement Actually Needed?
Third, and it is the previous page in one line: climb the ladder before you reach for the meter.
Does a search find it, because the answer already exists somewhere? Does a rule decide it, because the decision is a condition you could write down? Does a formula compute it, because it is arithmetic? If any rung holds, stop there — free, instant, identical every time, and auditable next March.
The model earns its meter on rung four, where language, judgement or reshaping is the work itself. Sorting a grumbling-but-warm review. Turning notes into a paragraph. Explaining a clause in plain words. That is not a small territory; it is most of what an office does with text. But it is not everything, and Q3 is where the difference gets noticed.
Q4: Is the Material Allowed In?
Fourth and last: Chapter 6, standing in the doorway.
Credentials never. Personal data about real people redacted, or not at all. Confidential material only where the tool is approved for it and the promise you made about that material still holds. These are not weighed against convenience; they are the categories that stay out regardless.
Which raises the obvious objection: if Q4 can kill anything, why is it last? Two reasons, and both are deliberate.
The first is that Q1 to Q3 are judgements and Q4 is a rule. Judgements get weighed against each other, and after weighing three things in a row the mind is warmed up and ready to weigh a fourth. Q4 does not weigh. Putting it at the end, alone, is what stops it being negotiated with at ten to five.
The second is that you do not know exactly what would go into the box until the first three questions have shaped the task. Q4 is asked about the material you are actually about to paste, which is a question you can only answer late.
And a task that passed the first three and dies here is not the card failing. It is the card working — usually into "redact this and it passes", which is a better outcome than either shipping it or abandoning it.
The Card
All four together, in the order they are asked:
One. What does being wrong cost? Cheap, embarrassing, or serious — and the answer sets how hard the rest of the card has to work.
Two. Can I check it, affordably? A fact in two minutes, a sample I can read, or a check that runs by itself. If none of those fits and the stakes are not cheap, the answer is no.
Three. Is judgement actually needed? Search, rule, formula first. The meter is for work that needs language or judgement.
Four. Is the material allowed in? Credentials never, personal data redacted, confidential material only where it is permitted to go.
Thirty seconds, and most tasks take ten because the answers are obvious. That is the same argument a pilot makes about the walk around the aircraft before a routine flight: the ritual is not there because flying is usually dangerous. It is there because the rare miss is expensive enough to justify the same thirty seconds every single time, including — especially — on the flights where nothing is wrong. Leave the aircraft there and keep the ritual: all four questions, and especially on the tasks where you are already certain of three.
Three Runs Through It
The itinerary draft. A week in the lake country for a couple celebrating an anniversary, and Tessa wants a first draft of the day-by-day. Q1: client-facing, so the middle lane. Q2: yes — the load-bearing facts are the places she names, and each is two minutes on its own website, which is exactly the check that caught the Harbourview Annex. Q3: yes, this is language and judgement from start to finish. Q4: the guests become "a couple celebrating an anniversary" and their booking reference stays out. Four passes, ten seconds, and she starts drafting.
The expense reconciliation. Four rows of supplier costs against what the booking system recorded. Q1: money, so it is not the cheap lane. Q2: checkable, certainly — she would have to redo the sums to check them, which is a warning already. Q3: is judgement needed? No. It is arithmetic with a condition on it, which is rung three, and a spreadsheet does it exactly, instantly, and identically forever. The card ends here, and it ends here fast — this is the most common way a task falls off, and it never even reaches the privacy question.
The client-list segmentation. Marketing wants the guest list sorted into who might book a walking tour next spring. Q1: internal planning, low stakes. Q2: yes, a sample of the sorted list is readable in twenty minutes. Q3: yes, that sorting is judgement about text — genuine rung-four work. Three green lights, and then Q4: the material is a file of real people with names, addresses and booking histories, and that is the category that does not get weighed. Not "no", though. Strip the names and the references, keep the trip histories and the notes that describe behaviour rather than identity, and re-enter at Q4 with material that is allowed in. Two minutes of redaction, and the task that was dead is running.
Which is what the card is for. Not to make Tessa cautious — she is already the most confident user in the building — but to make her decisions the same on a Tuesday afternoon as they were on a Monday morning. Consistency in the human is what the machine cannot supply, and it is the subject of the last page in this chapter.
- "A four-question framework is overkill for everyday asks." The whole card is thirty seconds and the everyday asks clear it in ten. Its value is not the four hundred tasks it waves through; it is the one on a Tuesday that quietly deserved a no.
- "Q1 said low stakes, so I can skip the rest." A low-stakes task still meets Q3, which asks why you are paying the meter for something a rule would do, and always meets Q4, because permission has nothing to do with stakes. The card is short precisely because every question survived.
- "Dying at Q4 means the task cannot use AI." It usually means the material cannot go in as it stands. Redaction turns most Q4 failures into passes, which is why Chapter 6 taught it as a technique rather than a prohibition.
- "The card decides for me." It structures the decision; your judgement still fills in every answer. What counts as serious, what check is affordable, what the material really is — the card asks, you answer, and that division is the point.
- This is the course compressed into an artifact: four questions, each one backed by chapters you have read, and teachable to a colleague in five minutes. Everything else in this book is the reason the card is right; the card is the part you carry.
- It is also what turns a capable user into the person a workplace can standardize on. Anyone can be the one who uses the model most. The card makes Tessa the one who can explain, out loud and the same way every time, why a task did or did not go near it.
Knowledge Check
What are the card's four questions, in order?
- Allowed in? · Judgement needed? · Checkable? · Wrong costs?
- Wrong costs? · Checkable? · Judgement needed? · Allowed in?
- Wrong costs? · Judgement needed? · Allowed in? · Checkable?
- Judgement needed? · Wrong costs? · Allowed in? · Checkable?
Why is the permission question asked last rather than first?
- Because it matters less than the other three questions do
- Because the workplace rules on it change too often to rely on
- Because it is a rule, and rules are not weighed against convenience
- Because it is only a formality once the first three have been answered
Tessa's expense reconciliation dies at Q3. Why?
- A formula computes it exactly, with no judgement
- The material is private, so it cannot go into the box
- The stakes are too high for the model to be involved
- There is no way to check the answer it would give
The client-list segmentation passes the first three questions and fails the fourth. What does the page say happens next?
- The task is dropped, since three passes cannot outvote a gate
- The task goes ahead, because most of the card was satisfied
- The task waits until a manager approves sending the list
- The names and references come out, and the task re-enters at Q4
You got correct