Reading Claude's Conclusions

Claude's Conclusions reads the account this course has walked through and writes what it finds — a headline, a short summary, and a handful of findings, each one telling Marleigh & Rowe what matters and why. Every figure in it is checked before it is ever shown, the same discipline every lesson in this course has held to by hand: no number reaches the page that the underlying data cannot stand behind.

Facts with names, not digits with a story attached

The model is never handed a blank page and asked to remember numbers. It is given a set of named facts — this account’s revenue, its contribution, a product’s own margin — and told to write a placeholder wherever a figure belongs, in the shape {{fact.key}}, rather than the figure itself. This account’s own revenue for this window, for instance, would be handed over as a fact worth $26,907.89, and its contribution as a fact worth $11,531.39 — the exact same two figures lesson 1 and lesson 4 already showed you. The model never sees those dollar amounts as text it could mistype; it only ever sees the keys, and the application fills in the real number afterward. That is the same rule this course's own lessons follow: nothing above this paragraph was typed either — every figure came from the account snapshot the same way.

What gets thrown away

Before anything from the model is stored, it is checked. Every placeholder must name a fact that was actually given — a citation to a fact that does not exist fails the whole sentence. And no figure may be typed by hand at all: not a dollar amount, not a percentage, not a bare count like a number of days. A sentence that types a figure instead of citing one is discarded in full, not shown with an asterisk or a caveat. Product names are the one deliberate exception, since a title like a five-pack of something is a name a seller would recognise, not a number the model invented.

What gets thrown away is counted, and that count is shown on the panel itself. A checker that quietly dropped half of what it was given and said nothing would be no better than no checker at all.

The failure mode is a quieter panel, never a confident wrong number

A finding whose figures cannot be verified does not reach Marleigh & Rowe’s screen in some softened form — it does not reach it at all. That is the trade this whole course has made from lesson 1 onward, in a different form each time: an unknown payout, a missing landed cost, a row that cannot be judged against the bar, are each left out rather than guessed at. Claude's Conclusions is held to the identical standard, by code rather than by a writer's care, which is what lets the figures inside it be read with the same confidence as the tables beside them.

You choose how it talks to you

Under Settings → How Claude writes to you there are two choices, and they apply everywhere — Claude's Conclusions, the advertising report and Ask Claude all read in the same voice, rather than a different one per panel.

Reading level — from “explain it like I’m 10” through beginner and intermediate to expert. This is not about how clever anybody is. Somebody can run a business for twenty years and have never once needed the word ACoS; writing to that is service, not condescension. Pick the level that matches how much Amazon jargon you want to be handed without explanation.

Tone — ten of them, from professional and matter-of-fact through friendly, caring and enthusiastic, to funny, sarcastic, Gen Z, and one called S.L.J. that swears at your fees without holding anything back. It is off unless you choose it, and it aims the language at Amazon and never at you. If a shared inbox receives your alert emails, that is the one to think about first.

Here is the part that matters, and it is the same promise the rest of this lesson just made: a tone changes how a report reads and never what it says. No voice may soften a bad figure or inflate a good one, and every figure in every tone is still a {{fact.key}} citation checked by the same code. The gentlest wording and the angriest are reading the identical numbers.

One practical note. The louder the voice, the likelier the model is to reach for “about forty grand” instead of citing the fact — and that sentence is thrown away, silently, exactly as this lesson described. If a report suddenly reads thin after you change the tone, that is what happened. The count of what was dropped is on the panel; a quieter report is the machinery working, not failing.

Changing the setting affects the next report written. Reports already on the page were written in the old voice and are left exactly as they were — the same reason a figure from March still reads as it did in March.

Where to find it

Claude's Conclusions appears on the Overview panel for every account, and its own longer form on the Advertising page. Both are part of the plan tier that carries + Claude after its name — a seller paying for that tier is paying for exactly the reading this lesson just described, not a different one.