The AI writes the review.The code does the maths.

The consultant works with small-business owners: he goes through their numbers and runs offline workshops built on training cases. He needed a prototype he could show to partners and test offline. The business owner goes through the diagnostics alone, and the consultant checks the result and releases it after payment.

There was no ready specification for the method. There were principles, cases from his practice, the format of his offline cases and his express-diagnostics spreadsheet. From these we built rules of the form “condition in the answers → point of attention → what to check”. Each rule has a priority, in the order the consultant set: first who makes decisions and controls the process, then stock and purchasing, then margin, and only then demand. For every rule, the consultant’s source material is recorded.

One more condition of his: never blame an employee when the failure lies in how decisions are made. So the system suggests training only when the employee has the authority and the checks in place and lacks knowledge specifically.

The easy route would have been to hand the answers to a large model and ask for a review. The text would sound convincing, but the model could recalculate the margin its own way, mix up periods or name a cause for falling revenue that the client never mentioned. The consultant would then have to check every line.

So the code does the calculations: average order value, conversion, margin from mark-up or cost of goods, logistics as a share of purchasing, return on advertising, payback on an investment, the month’s result after fixed costs and loan payments, and return on the business by the consultant’s own formula. 17 formulas in all. If a formula lacks inputs it is not run, and the review shows that. A gap never turns into zero.

The method’s rules pick the main point of attention, and when data is thin the system asks follow-up questions. The model gets ready-made section drafts, numbers in Russian format, the owner’s goal and the industry, and returns the text as separate fields in a fixed schema.

After the model, the code checks the text. A section fails if it contains a number that is not in the client’s answers or the calculations, a reference to an unknown fact, a causal link such as “because”, “so” or “led to”, or a claim of growth or decline without two comparable periods. There are seven such reasons in all. A failed section is replaced by a template the code builds.

The consultant’s notes have a filter of their own: phrases that promise a result, such as “you’ll get your profit back” or “without new investment”, are dropped. A review may say a step can reduce losses, but it does not promise profit.

The writing model was chosen on 12 October: eight models were compared on seven scenarios, and Claude Sonnet 5.5 won. If the model is unavailable or the spending limit is reached, the review is still built, with template wording.

When the client submits a request, the system freezes the result. The answers, calculations, chosen conclusion and checked text go into an immutable snapshot with a hash and the builder version number. The review page, the PDF and the editable slide deck are all built from that snapshot, so the numbers match in all three.

If the client changes the answers, a new version appears and the old link still opens the old one. A consultant’s rewording also creates a new version that has to be confirmed, while the numbers and analysis stay the same. The builder has gone through 20 versions, and reviews built by any of them open exactly as they were released.

The paid review holds the main conclusion, a “how the numbers changed” block (before → now), the first step, a practical block with a table that can be filled in on the page and downloaded for Excel, a 90-day plan and a training module of five cases in the order the consultant uses in his offline workshops. There are seven modules, chosen by the main problem, 35 cases in all.

The substance of the reviews was checked with runs on the test server: typical clients went from quiz to PDF and slide deck. They included retail with falling revenue, a client who answered “I don’t know” to almost everything, growing services, manufacturing with repeated errors, and rising revenue with a falling margin, eight types in all. The slides were rendered and looked at.

The prototype was built for showing to partners and for offline tests. The consultant himself moved some ideas to later stages.

One question about the method is still open. A 20–30% drop in revenue gives way, as the main conclusion, to problems with authority or advertising that does not pay back, because that is how the priorities are set. I left the ranking as it is: it follows the rules, and changing them is the consultant’s call. The training block is rare too: it needs a knowledge gap with authority and checks in place, and the client has to want to practise the skill.