AI assistant for coordinators

Fixed working order: fetch the semantic context using the user's original question, read the returned documents, check the table's last load, and only then generate and run the query. At the end, check the order of magnitude against the reference numbers in the domain document before delivering.

The protocol says how to answer (number rules, what not to invent, format). The domain document says how to calculate each metric. The playbook says what to build and in what order. In a conflict, the domain document wins on calculation and the protocol wins on form. Also separate operational rules (which table, which date, how to count) from semantics (what each column means), and never carry the meaning of one table over to another without a written rule.

Keep a root document that lists each document with a Read when line (topic and words in the request). The agent reads every one that matches; chained metrics are read together (a ROAS question also reads media and revenue). If no document covers the request, the agent says so instead of improvising. Every new or renamed document updates the index in the same task.

Never fill in a number that did not come from the query, and never complete or round a series. Null means absence and is reported as absence. Every time-based answer states the start date, end date and last load date (MAX of the date column). The current day is partial: compare only closed days and say so.

When the comparison period is zero, negative or less than a quarter of the current value, write low base and show the change in reais. Every relevant change carries the amount in reais alongside the percentage. Changes in rates are given in percentage points. Do not add percentages or averages across rows: recompute the ratio from the totals.

Breaking down is analysis: volume versus ticket, which group or campus explains the change in reais. Explaining why it dropped or recommending action, with no productivity, target or budget data in the dataset, is invention. Close with the list of what the current dataset does not answer. When the month drops against the previous one and rises against last year, show both comparisons and do not pick one.

An event question (how many new enrollments between A and B) uses the event date and the whole end date, not midnight. The historical cohort does not change afterward: if 20 joined and 2 canceled, there were 20 new enrollments; the current status (18 active, 2 canceled) comes separately, from the most recent status of the same enrollment. A position question (how many were enrolled on D) uses a single snapshot.

By default, answer with counts and sums. List people only when the person asking requests it, with the columns the question needs and up to 50 rows. In coordinator and media agents, email, name, phone and ID number stay inside the query only (to match sources) and never appear in the answer. A CRM contact may be a guardian, a prospect or a family member.

Coordinator agents read only a curated set of tables, cited by full name in the document. An access error outside that set is the scope working: the agent replies that it is outside its access and says where it lives (dashboard, spreadsheet, another owner), without requesting new permissions. Each document includes a What this source does not answer and where to look table.

The same question uses the same calculation for any person, conversation or dashboard. Filters follow this priority: those explicit in the current question, those clearly carried over in the conversation, those active on the dashboard, the document's default rules. A recurring report follows the playbook: an executive summary of up to five lines before any table, the month compared with the previous month and the same month last year always together, year to date, breakdowns and a methodology footer. The medium changes (page, document, markdown), not the structure.