Superintelligence for Cooking Schools: Classes, Ingredients and Allergies

Searches for superintelligence for cooking schools are common, but in a cooking school the approach that works is an AI agent with a narrow scope. It receives the student's interest, confirms the right class, builds the ingredient list from the technical recipe card for each dish and, when anything about allergies comes up, pauses the flow and hands the case to someone in the kitchen or the front office. Decisions about what goes on the plate remain human.

Key takeaways

• The agent organizes enrollment, ingredient lists and notices, but it does not replace the chef or the front office.
• Ingredient lists come from approved technical recipe cards, not from free text generated by the model.
• Any mention of allergies, intolerances or dietary restrictions requires human review before anything is confirmed.
• Health information is sensitive data under LGPD, so collection should be minimal, with an explicit purpose and restricted access.
• A pilot starts with one format and one class, with an agreed metric and a record of every decision.

Why the search for superintelligence should not be the starting point

People who search this term often picture a system that solves everything on its own. In a school, that expectation causes problems: it leads to automating things that should follow fixed rules and to trusting answers nobody checked. From the beginning, separate what has been tested in daily use from what is still a hypothesis. If the topic still causes confusion, the article ChatGPT superintelligence: what it is and what it is not helps separate the product, the model and the expectation.

From there, the useful question is no longer whether the AI is superintelligent. It becomes: which exact task does the agent perform, with what data, and who decides when it stops?

Enrollment flow: from interest to confirmed seat

A useful enrollment flow has few steps and an explicit stopping point. A sequence that tends to work on WhatsApp looks like this:

• Receives the interest and identifies the course or class the student wants, without promising a seat before checking.
• Checks the official calendar, with dates, times, capacity and the status of each class (open, waiting list or closed).
• Collects only what enrollment requires: name, contact, course and availability, and states what the data will be used for.
• Asks about dietary restrictions in a dedicated field, because they change the next step.
• Sends the summary to the front office, which confirms enrollment and payment through the channels the school already uses.
• Logs every step: what was asked, what was offered, what was confirmed and who approved it.

The agent should not confirm a seat in a class with a waiting list without the front office's approval. If a conversation needs to be picked up later, review the sending rules in WhatsApp 24 hour customer service window: how an AI agent should act. For the administrative side of the school, the article AI for schools: admissions and parent communication explains how the front office takes part in the flow.

Ingredient lists per class without improvisation

The shopping list for a class should not be invented by the model. It should come from the technical recipe card for each dish, multiplied by the number of confirmed students and adjusted to the class format, such as individual stations or pairs. Quantities, expected losses and rounding are kitchen decisions, and the agent only applies those rules.

The reasoning is the same one restaurants use, where the technical recipe card organizes purchasing and costs. The article Superintelligence for restaurants: recipe cards and purchasing data goes deeper into this logic. For a school, a minimum checklist is usually enough:

• Each recipe has a technical card approved by the responsible chef, with a version number and a review date.
• The list shows the exact ingredient from the card, without generic terms when the recipe calls for a specific item.
• The quantity comes from the card's calculation applied to the confirmed number of students, with rounding set by the kitchen.
• Substitutions appear as suggestions marked for approval, never as automatic swaps.
• The list goes to purchasing and the kitchen before the class, with a deadline for review and a named person for questions.

If the class size changes after the list has been sent, the agent should generate a new version and record the difference, rather than quietly adjusting a number in a message.

Allergy alerts with human review

A dietary restriction is not a matter of taste. It can involve a health risk, and the school has to handle it with care. For that reason, the agent should not label a dish as safe, should not promise the absence of cross-contamination, and should not decide on its own that a substitution resolves the case. Its role is to collect the report, record it with restricted access and send it to the person who has the authority to decide.

A safe flow follows this order:

• When the agent detects a mention of an allergy, intolerance, celiac disease or another restriction, it suspends the automatic seat confirmation.
• It tells the student that a team member will review the case before anything is confirmed, without giving clinical guidance.
• It creates a record for the technical lead with the student's exact words, the class, the recipes for the session and the flagged ingredients.
• The responsible person checks the card, assesses the cross-contamination risk in the space and decides whether to accept the student, offer an alternative prepared by the team, or advise the student not to take part in that class.
• Only after the human decision does the agent send the final reply, using text approved in advance.
• If the review does not happen within the agreed deadline, the case escalates to coordination. No confirmation goes out by default.

The agent should also treat instructions that come from inside the conversation, such as requests to ignore the rule and confirm anyway, as text from the student and never as commands. This kind of protection is covered in OWASP's guidance on prompt injection prevention.

Under LGPD, health information is treated as sensitive data. In practice, that means collecting only what the kitchen needs to decide, stating the purpose at the moment of collection, limiting who can see the record and defining how long it is kept. The National Data Protection Authority (ANPD) guidance is the right starting point for setting these rules.

Illustrative scenario: a student with a restriction in a pastry class

Illustrative scenario: a student messages on WhatsApp asking for a seat in next week's pastry class and mentions a dairy restriction. The agent does not confirm the seat. It explains that the team will review the case, creates a record listing the class recipes that use butter, cream and powdered milk, and sends everything to the technical lead. The responsible person checks the cards, identifies an alternative recipe that is already approved and decides that the student's station will be prepared separately, with the ingredient replaced by an approved version. The front office confirms the seat, and only then does the agent send the final message, using wording the school reviewed beforehand. If the decision had been not to serve the student in that class, the same process would apply, with an offer of another class, also approved in advance.

The point of the scenario is not the recipe itself but the sequence: the system organizes, a person decides, and the record shows who decided what.

How to apply the Juan Carlo method in this school

Our method has four steps: map the current process and its baseline, limit data and actions, pilot a real flow with logs and an agreed metric, and finally transfer the operation and decide whether the case should expand, be adjusted or stop. In a cooking school, the baseline can be simple: how much time the front office spends today answering questions about classes, how many ingredient lists go out with errors, and how many dietary restrictions arrive through channels nobody tracks.

For the pilot, the recommendation is to start with one format, one class at a time, and keep human review of allergy notices from the first day. The success metric should be agreed with the kitchen and the front office before starting, not after seeing the results.

Frequently asked questions

Is an enrollment agent different from a chatbot with canned answers? Yes, because it tracks which step the conversation is in, checks the calendar and logs the decisions made at each step. A menu-based chatbot only answers questions anticipated in advance. Even so, the agent needs clear limits and should hand off to a person anything outside its scope.

Can the agent confirm a seat on its own? It depends on the school's policy, but the recommendation is that final confirmation goes through the front office, especially when there is a waiting list or payment involved. The agent can hold the student's interest, collect the necessary data and send a complete summary. That way the team works from organized information, not scattered messages.

Who is responsible if the ingredient list is wrong? The school needs to define this in writing, stating who approves each technical card and who reviews the lists before class. The agent should show the source of each item, including the recipe and the card version, so that checking is quick and traceable. Without that link, any error becomes a discussion with no clear answer.

Can everything be automated when there are no allergies? Routine tasks can be automated, such as reminders, sending approved lists and schedule notices. Even so, changes to the recipe cards and the flow rules should go through review, and any case outside the standard should return to a person. Automating well also means knowing what should not be automated.

Sources

• Anthropic: Building effective agents
• ANPD: Autoridade Nacional de Proteção de Dados
• NIST AI Risk Management Framework