Turn Monday's exports into a finished report draft

The assembly and the first draft of any recurring report built from exports: CRM pipeline, ad spend, revenue, project status, whatever your Monday numbers are. The assistant fills your template, compares each metric to last week, and drafts the narrative — with hard rules that force it to flag missing or conflicting numbers instead of papering over them. Anything it can't verify gets a visible [CHECK] tag, not a confident guess.

What it doesn't handle: pulling the data itself. You still export the files each week. That's deliberate — a human pulling saved views is the simplest reliable data layer you can run without building anything.

Setup happens once. You write a report spec (every metric, its source, its comparison), save an export view in each source system so the weekly pull is identical every time, and use the included prompt to turn your best past report into a reusable template.

The weekly run is four moves: export the saved views, paste them into a fresh chat with last week's report and the template, run the assembly prompt, then work through the review list it produces. The recipe ends with a sanity-check list (spot-check three numbers, confirm date ranges match, resolve every flag) because the draft is only as good as the check you run on it.

Honest limit: this recipe works while your sources are few, the exports are clean, and the report stays internal. It breaks quietly — a system changes its export format, a date filter drifts, and the report is wrong for three weeks before anyone notices. If the report goes to clients or drives real decisions, the monitoring is the actual product: an agent that pulls directly from your systems, checks totals against the source, and escalates anomalies. That's the done-for-you version, and a free audit will tell you whether your reporting justifies it.